Welcome to the Anime Diffusion Thread. This is a thread for discussion relating to leveraging generative AI technologies for anime-related content. Images, videos and sound all welcome. Cloud-based anime generation is also welcome.

Please keep all non-anime gens on >>>/ldg/ >>>/sdg/

Because this is a new general, we are going to borrow a lot of the /ldg/ OP until we get some more anime-focused guides.

>Generating Images
https://rentry.org/comfyui_guide_1girl
https://tagexplorer.github.io/

>Generating Videos
Guide: https://rentry.org/wan21kjguide
https://github.com/Wan-Video/Wan2.1

Oldfag's wise experience: >>106082956 >>106082967

>[NEW MODELS!:]
Name: ItN's Vintage Merge
URL: https://civitai.com/models/1823418/itns-vintage-merge

Anime Pop Cute Illustrious-XL
https://civitai.com/models/1540768/anime-pop-cute-illustrious-xl

Name: Retirement Mix (NAIXL)
URL: https://civitai.com/models/1821991/retirement-mix-naixl

>[UPDATES!:]
Name: Smooth Mix (Illustrious) v5
URL: https://civitai.com/models/1695233?modelVersionId=2038033

Name: Hotaru Blend
URL: https://civitai.com/models/1696665/hotaru-blend

Name: NanoMix ~Multiverse Remix~
URL: https://civitai.com/models/932505/nanomix-multiverse-remix

Name: Mature ritual
URL: https://civitai.com/models/994401/mature-ritual-or-illustrious

>Anime Checkpoints and loras for Flux and SD3.5
[SD3.5] Anime Style LoKr - Large / Medium
https://civitai.com/models/1807230/sd35-anime-style-lokr-large-medium

[FLUX] Illustrious Anime Style Lora/Lokr
https://civitai.com/models/1059388/flux-illustrious-anime-style-loralokr

[FLUX] 1984 - Anime Style feat. IllustriousXL
https://civitai.com/models/1072833/flux-1984-anime-style-feat-illustriousxl

Anime3.5L
https://civitai.com/models/931532/anime35l

Previous: >>106051736

>>106087463 (OP)
fern is sex

Okay, now stop the damn joke.
Please, nobody post in this thread.
Let it be.
Let it die.

>>106087497
Kys tranny. Shouldn't have ran the site to the ground to the point only people that hate each other remain and they split into 100 dead generals

SUMMARY OF LAST THREAD! 1/3

Resources shared (highlights)
- Guides: ComfyUI 1girl guide, Wan 2.1/2.2 video guide and repos, rentry video workflows.
- Anime models/checkpoints/LoRAs: Illustrious/SDXL-focused anime finetunes (Illustrious Anime Style, 1984 feat. IllustriousXL, Anime3.5L, SD 3.5 Large Modern Anime, flux.1 dev modern anime, WAI NSFW Illustrious SDXL, NoobAI XL, etc.), plus a fresh merge “LumiNetaMix.”
- Workflow/asset hubs: civitai, tensor.art, openmodeldb, openart workflows, civitaiarchive.
- Tagging/tools: WD tagger, character select SAA, NAIfags wikis, aibooru, nax.moe.

General discussion themes
- Composition/distance issues: suggestions included “long shot,” Zoom Slider LoRA, and prompt structuring.
- Adetailer blending issues: adjust detection threshold, mask padding, denoise, sampler consistency; anime face segment model; keep denoise lower to avoid “boxy” blends.
- LoRA training tips: 20–30+ varied images, SDXL at 1024 or 768, clean tags, avoid over-adjectives, lrs ~1e-4, lower ranks (e.g., 8), optimizers (Lion/D-Adapt), test epochs.
- Model merging: You can’t import “only good” traits—merging mixes weights and can bring artifacts; block merging helps on SD1.5 but less reliable on highly stylized SDXL models; better to retrain characters on the higher-quality base.
- BREAK prompting (Rouwei): Several users note Rouwei benefits from BREAK-chunked prompts (e.g., character | clothing | setting | pose) for coherence; others report it’s usable without BREAK but performs best with it.
- Prompting style split: Tags vs natural language. Some users praise newer models (e.g., Neta Lumina) for handling NL prompts decently, but most still recommend Booru-style tags for control and adherence.
- Inpainting: 1girl-focused checkpoints are bad at general inpainting—use a more general SDXL/SD3.5 base for inpaint, mask larger regions (whole head), use 0.4–0.6 denoise, and inpaint characters separately in multi-char scenes.

>>106087592

2/3

- Upscaling: Base gen at native res, then upscale 1.5x–2x with good anime upscalers (4x-AnimeSharp, 4x-Nomos8kDAT), use tiled VAE to save VRAM, then detail pass (Ultimate SD Upscale + detail LoRA).
- Hardware: 16GB+ VRAM preferred; use tiled VAE; A1111 seen as less VRAM-efficient; Forge/Re-Forge or ComfyUI recommended.
- Flux/Chroma: Chroma (Flux-based) looks promising but LoRA training and finetunes are harder; Flux anime finetunes are emerging but not as mature as SDXL anime.
- ControlNets for SDXL: Mixed results vs SD1.5; require proper SDXL CN models; updates in Comfy often needed.
- CivitAI/region issues: UK blocks CivitAI; users share alternates and archives.
Video models and results
- Wan 2.1 vs 2.2: 2.2 adds more data and anime but still mixed results for anime motion; 2.2 workflows in Comfy require two models (refiner pipeline). Some report better visuals, others complain about tweening, hand smears, extra limbs, weak prompt adherence at low CFG.
- AniSora/AniWan: AniSora praised for anime-like motion but heavily censored; AniWan/AniSora variants recommended for anime if you can get them. Wan 2.2 examples posted; performance varies.
- Q/GGUF variants: Some testing with Q6/Q8 for video; speed/quality tradeoffs discussed.

>>106087601
3/3
Model news and drops
- New/updated anime mixes listed: ItN’s Vintage Merge, Anime Pop Cute Illustrious-XL, Retirement Mix (NAIXL), Smooth Mix (Illustrious) v5, Hotaru Blend, NanoMix Multiverse Remix, Mature ritual, C0RRUPT.F33D.SDXL (creepy styles).
- Flux anime finetunes and merges trickling in (e.g., coamixflux). Interest in Illustrious 2.0 finetunes; waiting for 3.0 is seen as unrealistic due to Stardust gating.

Community/meta
- Thread adoption slow; some argue /g/ isn’t ideal for anime art discussions; suggestions to shill on /h/, /e/, /a/, or elsewhere; /ic/ would be hostile. Some friction with non-anime posts and “schizo” accusations.
- GenJam events and polls ongoing; some videos/albums shared.
- Brief tool questions: metadata viewers on mobile, tag completion in Comfy (ComfyUI-Custom-Scripts), AMD suitability (not deeply discussed), Comfy random soft-freezes (likely environment/VRAM/tiling/large workflows).

Practical mini-guide (from anons)
- Use Illustrious/NoobAI for anime; WAI/others for learning; avoid Pony for 2025 workflows unless you need its style.
- Tags > NL prompting for control; use Danbooru wiki; separate character/environment chunks; avoid overusing quality tags if it flattens output.
- For LoRA training: clean datasets, invented triggers, thorough tagging, many images, test multiple epochs.
- For inpainting: use general base, larger masks, mid denoise, consistent sampler.
- For high-res: base gen -> upscale -> detail pass, tiled VAE on.

LISTEN UP YOU FUCKING TRAITORS

I CAME HERE FROM TO TELL YOU PIECES OF SCHIZOS EXACTLY WHAT I THINK OF YOUR LITTLE ""THREAD""

You think youre so special??? You think anime deserves its own thread???? WE'VE BEEN DOING ANIME IN /ldg/ SINCE DAY ONE!!!!

You know what you've done??? You've FRACTURED the community!! Now newfags dont know where to post! Now we have to check TWO threads for updates! CONGRATS RETARDS YOU PLAYED YOURSELVES

>>106087612
Which is the ongoing GenJam?

>>106087774
okie dokie :D
>>106089172
sounds expensive

>>106087774
>>106087780
Surely we can reach a compromise, how about spoilering all 3dpd?

>>106089864
windoze

>>106089999

cute thread, CUTE! :3

<3

>Hm, maybe this general doesn't have postcard in it
>It's even worse here

until tomorrow <33

snowflake general

>>106091501
more flies with honey than vinegar larpy

>>106090522
havent seen an anime this bad since i downloaded crunchyroll
beahagahaha

>>106087780
>>106087774
BEAHAGAHAH

>>106089999
>>106090193
Nice butt.

>>106087463 (OP)
ENOUGH WITH THE AI DIFFUSE MY COCK IN YOUR MOUTH GENERALS YOU FUCKING FUCK

>>106091626
Look at him. Look at him and laugh.

>>106087463 (OP)
hot damn that's just what the doctor ordered, a nice fat pair of anime tiddies

>1girl

>>106093234
>2girls

>>106092444
>>106092555
checked

how is this general /g/ related?

>>106094721
because your /g/ay

>50 posts
>nor a post that talks about anime difussion

SDXL is still king.

a-anime diffusion?

>>106094359
>3girls

>>106095628
how far can we go

>>106095628
Do you have a workflow togenerate qith wan?

>>106095647
>4girls?!

>>106095628
>>106095776
Why is it so blurry?

>>106095776
moar

>>106095776
4 GIRLS 4 FROGS

>>106094721
the images are created using various open-source image diffusion software.
just as on topic as any other Ai general.
>>106095776
cute, CUTE!

until tomorrow<3

Wan 2.2 workflow for 1 frame images?

>>106099691
salvage frames & edit

Can we organize and promote this general in other boards? We must work together if we want /adt/ to take shape. This is what I didn't imagine at first!

I saw the possibility of creating a general where we could share, in addition to gens, information about our experience with anime genning, comments on models and Loras, from AI artists to other AI artists.

ANOTHER ONE? come the fuck on I have adhd every time I have to switch tabs my short term memory is wiped I can't keep doing this

>>106091038
dif posts/styles in each general, seems like a flex

>>106087463 (OP)
Can I generate front and side views of a character with proper proportions so it can be used as a reference for 3d modelling?

>>106101039
Yes, just download ComfyUI and IllustriousXL

>>106101066
thanks, fren. I'll give it shot.

>>106087463 (OP)
fern is made for old rich bald men

>>106101638
haha cuuute

>>106102064
"the sun will never rise again..." - Sakurasou No Pet Na Kanojo

is that her?
have i done it?
become the weeb?

>>106102401
Yes, it's her.

>>106102460
ok so i NEVER saw the show
& was able to refind the name based on the quote i provided

i just like that specific artstyle
& had it set as my desktop background for months kek

>>106102542
edited for cameltoe ;3

Why do you have so many threads and what's the difference?
>edg
>hdg
>ddg
>this one
>ldg
>probably something anime-related in /trash/ as well

>>106102707
so many threads? so much more FUN
there are many threads discussing the topic of Ai, for dare i say it, FUN! ;3

>>106102707
if you are just now getting into Ai
& have a decent GPU visit \ldg\
+ learn how to create LOCAL
then they can't tell you how
to play the game...
its almost over
<3

>>106102707
>edg
NSFW erotic, no sex.
>hdg
NSFW hentai, sex, with males involved (fucking the girls)
>ddg
Extreme and uncommon fetishes that aren't welcome anywhere else.
>this one
Safe work work anime images
>ldg
Video spam, "tech" discussion and 3dpd shit
>probably something anime-related in /trash/ as well
Like what?

>>106101638
did you put the censor yourself or the cuckedAImodel?

>>106103197
nta but he definitely did it himself

ehat wnt wrong with this general?

>>106103520
noffin ;3

>>106102870
Salty kek

>>106104395
he's not wrong

>>106104411
No he's just a very salty boi

>>106104429
salty but correct

>>106101638
post the uncensored one somewhere anon-senpai

>>106105711
fuck off just see hentai or something LOL

>>106102870
There used to be a thread on /c/ for a while that would have worked for the purpose of this thread, but it's dead.

>>106105830
don't be a cunt

>>106105934
he's german so that's a difficult request

this is the worst thread of my life

epic necrobump

>>106106411
you should see /hdg/ or /aids/

why we dont organize and promote this general, it can be a good one, but you always leave me allone with this.

>>106106458
Are you the OP that literally shat himself publicly because no one gave you (You)s for your slop?

Someone is baking?

>>106106497
im not the OP, im just an anime sloper who doesnt have his dream general

>>106106571
Are you scared of realistic girls or what is the point? Not judging just trying to understand the splitting mentality

>>106106571
i want a thread that discuss AI generated anime, similar to discord, similar to 2ch or 5ch. This is an Anime website, why anyone is interested in this?

>>106106595
3DPD ("realistic girls") is pretty ugly compared to anime girls. From a purely aesthetic point of view, they are disqualified. Why would anyone want to look at ugly things?

>>106106615
No, we must organize and develop a long term plan. Simply image gens will only get us deeper into oblivion.

>>106106628
>From a purely aesthetic point of view, they are disqualified.
And from a purely spiritual, personality point of view they are discualified also. The only thing that sustains real women is that they are the only ones who are real (for now)

>thread full of virgins
Yeah, stay here please

>>106107371
you have to go back

>>106107432
What do you mean?

>>106107465
fuck back off to reddit or whereever normalfag site you came from

>>106107520
There's always the possibility of getting a prostitute anon
You don't have to live like this

>>106107520
>>106107656
you two could just fuck each other...

>>106107656
Those are not anime girls.

>>106107703
I'm not a homosexual but i might be in the mood of paying him a prostitute so he can experience being normal

>>106107751
>I'm not a homosexual
nobody will know your boipussy got stretched bro

>>106107751
I could pay for my own. I have a very well paying job and no issues interacting normally with people, you retarded faggot.

>>106107849
Sure thing buddy
Why don't you ahead and post your anime children while sitting alone in your dark room?

>>106107936
My room is well let and I will, tyvm. Now go back.

>>106108000
lit, fuck

Gens?

>>106087463 (OP)
Fern sexo

>>106087463 (OP)
we need an /ai/ board already and holy BOOOOOBS

grim

>>106109243
this, it should have been a futa bulge in OP

>>106106411
BEAHAGAHAHAHA

>>106087463 (OP)
nice slop you guys are making hier

>update
>no new posts

137 messages talking about nothing AI related
137 messages talking about nothing Anime related
137 messages talking about nothing Difussion related

>>106109766
r u a matematician or sumtn?

mh.

we have a nerd over here

>>106105711
Do you really need to see the uncensored version though? These characters are under the age of 18, therefore they are not of legal age.

>>106109980
https://www.youtube.com/watch?v=OMVc5R-cfS8

>>106097179

>>106110431
schizophrenic situation

>>106110860
your answer /adt/?

>>106110884
You are the only one in this thread, so you must represent /adt/ outside of this thread.

>>106110884
lonely gooning

>>106110884
>>106110630

>>106111002
see >>106107656

more flies with honey than vinegar anon

Why, why I have to goon alone?

>>>/a/

status?
>we sent a high level unit sir, code pink.

Alright, here’s the no-bullshit version from an old /adt/ anon:

Models

- novaAnimeXL_ilV90: decent with regional prompting, but easy to overconstrain and wash out if you cram too many attrs.
- Inpaint: use an actual inpaint model or ControlNet Inpaint. If editing a hand-drawn piece, match style (base+style ref), or you’ll fight seams forever.

LoRAs, tags, prompts
- ComfyUI: `<lora:Name:0.5>` in text doesn’t do shit unless you use parser nodes. Load LoRA via node; use trigger tag only in text.
- Synonyms mostly work; don’t be a slave to booru tags. But some tags nuke style: “young” often forces bean-mouth/round eyes. Know your model’s quirks.
- Character oddities (e.g., Ryuko’s pupils) come from data. Want normal? Negative it or override with a style LoRA.
- Adetailer: run on face, not just eyes. Eyes-only crops can shift colors. If it still derps, change its inpaint model and strip conflicting style from its sub-prompt. Bad LoRAs cause 90% of eye pain.

Inpaint/workflows
- Forge is snappier for quick inpaint. Krita plugin is only good if you build a clean Comfy custom workflow and expose minimal knobs. Otherwise, stick to your main UI.
- Big edits: use inpaint model + pose/depth ControlNet. Small fixes: rough it by hand, then inpaint to clean.

>>106111903
Regional prompting (Forge Couple)
- Don’t overstuff regions. Keep global scene in main prompt; per-region = identity + 1–2 actions/emotions. Lower CFG (5–6), saner sampler (DPM++ 2M SDE Karras/Euler).
- Too many regions? Merge where possible. Use pose/depth first, then layer style.
- Hires Fix crashing? Ditch Hires. Do two-pass: txt2img latent upscaler/ESRGAN light low-denoise/inpaint cleanup.

Upscale
- Extras can be mush. Use latent upscale or SDXL refiner; ESRGAN anime/4x-UltraSharp, then mild denoise pass to restore lines.

Captioning (for NSFW/datasets)
- Skip old JoyCaption unless you’re desperate. Local: QwenVL2.5-3B/Gemma-3-4B for fast captions; ERNIE 28B or bigger if you’ve got VRAM; Torii for NSFW/char names. Plenty of Comfy nodes and CLI tools exist.

Rules of thumb
- If eyes keep sucking: test without LoRAs. Train your own if you must.
- If adetailer shifts colors: clean sub-prompt, use a different inpaint model.
- Multi-girl scenes: layout first (Pose/Depth), style second, keep regions lean.
- Comfy is god-tier only if you build a sane graph. Fewer sliders, faster iterations.
- When in doubt: simplify sampler/CFG/prompt, add complexity back slowly.

>>106111903
>>106111913
Good stuff anon

>>106111101
is this WAN?

>>106113125
Yes. Wan2.1.
Wan2.2 is better

>>106113339
wish i knew how to do stuff like this that keeps it in the show's original artstyle.
i'm a noob and all i've ever been able to generate is the very generic glossy-looking stuff, no idea how to prompt specifically to get it like that

>>106113413
I recommend reading the guide in the OP:

https://rentry.org/comfyui_guide_1girl

Pay particular note to the part about controlnets.
For images that look just like a TV anime, I recommend using loras from civitai because danbooru doesn't have that many anime screenshots; certainly not enough to help generate any character in their anime style.

>>106113438
i appreciate the tips, i only started not long ago so there's a lot to figure out, but that guide looks great.
do you have any suggestions for loras? there's a ton on civitai and it's hard to sift through them all.
though, that may be a stupid question since all animes are going to have different artstyles, so it's not like there's going to be a one-size-fits-all.
i suppose the more important thing is figuring out what keywords to search for that will get the proper show-accurate lora artstyles, i'm still lost on all that.

>>106113576
You have to figure out what character you want, search civitai and tensor and judge by the thumbnail/preview what works for you. Always favor Illustrious over Pony.
Some loras are poorly trained and you might have to download alternatives if one isn't giving you the results you desire.

>>106110410
that's the reason I need to see them

>2.2 guide and workflow posted
spoonfeed me pls

>>106114492
https://rentry.org/wan22ldgguide

>>106114535
yeah I actually saw it, am searching for the quantized models now (16gb ramlet here)

>>106114542
https://huggingface.co/bullerwins/Wan2.2-I2V-A14B-GGUF/tree/main

>>106114542
i have 16gb vram and it works for me. Tho i also have 64gb dram which matters for swap

>>106114589
I also have 64gb of dram, I'll try both the fp8 and Q6 model to see if the speed tradeoff is worth it. with wan2.1 sometimes my gpu crashed tho, I blame shitty njudea drivers

>>106110410
The anime characters you mean? In this drawing they are all 35+, no worries mate.

>>106114596
try increasing blocks_to_swap

>>106111903
>>106111913
How do you get IL models to work with DPM++ 2M? Mine seems to keep just churning out blocky artifacts so i've been settling with euler for a while

>>106115071
oh nvm i think i found some potential solutions lol

>186 secs to gen with the fast 2.2 workflow
HOLY shit, it took at least 15-25 minutes on 2.1 (and Q6)
I'm using fp8 fast now, im gonna FUCKING GEN

gm guys

>>106115071
depends of the checkpoints, usualy i m using DPM++ 4M with Karras as denoiser, other schedulers don't work very well with DPM++ series.

For general I also lower de CFG when using DPM like 2-3, also DPM is a sensible sampler, i always use it as hiresfix sampler because add a lot of textures and details to it.

EXLUSIVE /adt/ WAN 2.2 INFO, 1/3
Release and models

WAN 2.2 released (open-source). Two-branch architecture: high-noise (motion/structure) and low-noise (detail/refine). 14B is actually MoE (~27B total), plus a smaller 5B all-in-one (T2V/I2V, 480p/720p).
New WAE for 5B only; 14B still uses WAN 2.1 VAE.
GGUF and quantized checkpoints are already appearing, but quality varies; many consider GGUF on high-noise possibly acceptable due to initial noise being “rough,” but trade-offs are real.

Performance, VRAM, and speed

14B 2.2 is heavier than 2.1 by ~2–3 GB VRAM. On 12 GB cards you’ll likely need offloading/block swap; on 8 GB, 5B fits with ComfyUI native offload (still slow).
Without accelerators, 14B 2.2 is slow (tens of seconds per step). With WAN 2.1 Lightx2v LoRA hack, people get 6-step (3+3) workable runs, but quality worse than native 2.1 + lightx2v. Proper 2.2 distill/LoRA is awaited.
torch.compile can speed some setups but may cause CUDA instability for some users.
SageAttention v2 increases RAM/VRAM use by ~25–35%. Some switch back to 1x attention or try radial attention (claimed ~40% speedup, quality TBD).

FPS, duration, interpolation

Base WAN remains effectively 16 FPS. 5B samples sometimes look like 30 FPS but are often interpolation-ish; real temporal frequency still bound by base. Many users interpolate to 24/30/60, but note: speed depends on final video FPS you set, not the interpolation itself. To keep motion speed unchanged: Final_FPS = Multiplier × 16 (e.g., target 32 or 64, then render to 30/60 respecting ratio to avoid slow-mo/fast-forward).
WAN 2.2 supports longer contexts like 121/145 frames (7.5–9 s), but compute cost rises nonlinearly.

>>106115708
2/3

Quality and behavior

WAN 2.2 follows prompts better than 2.1 (notably hands/arm crossing and complex actions), shows more expressive faces, and better camera motion realism. But outputs often feel noisier/less clean without tuned settings.
5B quality is clearly below 14B; good for “previewing” motion/idea and low-VRAM users.
Animation and anime styles often look better than live-action realism in early tests.
Some users see artifacts/“flashes,” unwanted camera wobble, underspecified objects; tuning samplers, noise branch steps, and weights helps.

Using high-noise/low-noise correctly

High-noise drives motion, layout, lighting; Low-noise refines detail. For I2V especially, high-noise is central; low-noise alone looks like WAN 2.1 detail pass. Most use a 2-stage sampler chain: first high-noise, then low-noise.
Typical quick test: total 5–6 steps (e.g., 2 HN + 3–4 LN) with an old Lightx2v LoRA to keep speed usable. Many report better results by increasing LoRA weight more on high-noise than on low-noise in 2.2.
add_noise must be enabled on both samplers to avoid washed/flat results in 2.2 base workflow (bugbear noted by users).
Shift/step-split tuning matters; users experiment with start_at/end_at distribution between HN/LN for style vs identity preservation. For I2V identity, consider fewer HN steps, stronger low-noise and maybe text weights lowered.

>>106115716
3/3


LoRAs and accelerators

WAN 2.1 Lightx2v (StepDistill/CfgDistill) “kind of” works on 2.2 but degrades quality; still valuable to keep runtimes sane (e.g., 6 steps total). Expect a dedicated 2.2 distill/LoRA later; community anticipates weeks to months.
Some community LoRAs: facefuck, motion LoRAs, and “fastvideo” were tested with mixed success. Weights for 2.2 often need to be higher than 2.1, especially on high-noise branch.
A few all-in-one “Rapid” mixes that blend 2.2 HN/LN + 2.1 + Lightx2v exist, but many call them low-quality hacks that cancel 2.2’s gains.

Workflows and tools

Official ComfyUI examples: https://comfyanonymous.github.io/ComfyUI_examples/wan22/. Some had errors until users cleaned or updated custom nodes. Safe practice: disable customs, test clean, then re-enable incrementally.
Kijai provided alternative ComfyUI nodes and block-swap workflows; also provides Lightx2v LoRAs and examples. His WAN 2.2 workflows support 81/145 frames and 6-step “accelerated” runs but quality is behind true 2.2 distill.
Users report 2.2 I2V workflows dropped CLIPVision nodes; consensus: it rarely helped anyway.

Practical tips

For speed, keep resolution/frame count modest; use block swap/offload; try old Lightx2v with caution. Don’t expect 1080p/9s at decent speed on 12 GB without aggressive compromises.
If CUDA randomly dies, try disabling torch.compile.
If outputs look washed: ensure add_noise on both samplers, tune CFG; consider euler/beta samplers (common combo).
For I2V identity preservation, push more steps/details into low-noise; for motion and dynamics, increase high-noise steps/LoRA weight.
Prompting: WAN 2.2 has an official prompt guide (AliDocs). Some found non-English (RU/CN) terms sometimes work better for body attributes. For “big ass,” try combinations like “wide hips,” “plump/bubble buttocks,” etc.

1/5
EXLUSIVE /adt/ WAN 2.1 INFO

Hardware, performance, and settings

>Typical speeds (optimized):

RTX 4060 16GB: ~2 minutes for 5s 480p with Lightx2V, 4–5 steps, blockswap 12–20.
RTX 4090: Commonly 1 minute or less for 5s 480p with Triton + SageAttention + Lightx2V.
RTX 2060/2060S: 15–20 minutes for 5s without accelerators; Lightx2V is essential to make it reasonable.
RTX 4070 Ti (12GB): Usable, but VRAM-limited for high res; still fast with Lightx2V.
RTX 3090 (24GB): Good balance; can run 720p decently and supports heavier workflows.
H100/A100: Not discussed with concrete numbers for WAN; overkill but fast in cloud.

>VRAM vs. chip:
Chip speed matters a lot. VRAM matters for resolution/context length/upscaling.
GGUF offload to RAM costs roughly 2× speed slowdown vs full VRAM, but still usable (not like LLMs).
32GB RAM is a minimum for comfort; 64GB is the sweet spot. Heavy workflows can blow past 32GB; swap helps but slows.
Blockswap vs. GGUF:
GGUF is easy but slower and can degrade quality vs safetensors. Blockswap lets you keep safetensors and only swap a few blocks to fit VRAM with minimal speed loss.
File swap/swap delays:
Large delays often come from reloading models/loras between runs. Kijai added nodes to reduce reloading, and native workflows help.
Linux vs Windows:
Linux can be finicky with fan curves and performance; manual scripts help. Windows is more plug-and-play.
CRF and video output:
Lower CRF = higher quality and bigger file. CRF=0 gives maximum quality (huge files). CRF=19 is a balanced default.

>>106115785
2/5
>Accelerators and installs

Triton (Windows builds): Use woct0rdho’s builds; install via pip with URLs. Sometimes need Visual Studio redists. For portable ComfyUI, install in python_embeded.
SageAttention:
v2 works best on Ampere/Ada/Hopper (RTX 30xx, 40xx, H100). RTX 20xx has limited/fragile support (some reported the old SA works, SA2 often not).
SA3 teased; optimized mainly for RTX 50xx at first.
Helper packs:
crossOS_acceleritor and helper_comfyUI_accel install Triton/Sage/etc. One-click-ish on Windows. People report big speedups (e.g., WAN 720p from 40 min to ~150s on RTX 30xx).
Torch compile: Works in native WAN workflows; be mindful of combos with Kijai nodes.

>>106115708
>>106115716
>>106115724
cheers anon

>>106115797
3/5
>Models, LoRAs, samplers

Lightx2V distill LoRAs:
New rank64 for T2V/I2V (official lightx2v repos) are strong. Rank64 is often optimal; rank128+ can hurt motion.
Kijai variants exist (rank64/128/256; adaptive quantile). Community reports: I2V rank64 is a “win”: more movement, less oily motion, more prompt sensitivity.
Some T2V Lightx2V rank64 had noise due to a bad release; fixes rolled, Kijai published alternatives.
Use matching LoRA type: T2V LoRA on I2V often causes slow-mo; vice versa too.
Strength: Many use 0.7–0.8 at 5 steps to avoid overcooking. 4 steps is very fast but can look artificial.
Other speed LoRAs:
PusaV1 (RaphaelLiu) + Kijai wrappers: mixed early impressions; feels different, sometimes weaker adherence; faster for T2V.
NAG:
Best used with native NAG ksampler (ComfyUI-NAG). Regular negative is ignored at cfg=1; feed “NAG-negative” properly. Gains are modest; can slow gen. Often better to weight positive tokens and skip complex negatives on cfg=1 Lightx2V runs.
Samplers:
LCM popular for fast 4–5 step WAN. UNI3C discussed, but it’s camera-control context, not a magical adherence boost.
FP8 notes:
fp8_e4m3fn_fast sometimes degrades quality; try non-fast. For RTX 30xx, use fp8_e5m2 weights if needed.

>>106115805
3.5/5

>I2V vs T2V, MagRef, SkyReels

I2V “MagRef”:
Great for face/identity retention; base image or batch works; minimal “face drifting.”
Use I2V-appropriate Lightx2V LoRA to avoid slow-mo/sticky motion. Works at 480p, can push to 512–656 square-ish. 720p possible with more VRAM.
SkyReels v2:
Cinematic faces/emotions; stylizes more. Supports genuine 24 fps sequences without slow-mo artifacts. Still early; no official distill yet. Good alternative for certain looks.
VACE:
Powerful for edit-style workflows (inpaint, cloth change, face swap compositing), context-based stitching. Heavy but flexible.
Slow-mo issues:
Caused by mismatched LoRA (T2V LoRA on I2V), or LoRAs trained on slow footage. Fix by using correct LoRA, mixing a fast-movement LoRA, or removing problematic ones.
i2v typically a bit slower than t2v at same res/steps, but close with optimizations.

>Upscaling and interpolation

Interpolation:
RIFE/FILM/GIMM VFI nodes work. Stick to simple: feed frames from VAE decode to RIFE VFI; set multiplier 230fps or 460fps (WAN native ~16fps). Then combine with matching output FPS to avoid speed change.
Third-party tools like FlowFrames can be simpler/faster out of Comfy.
Upscaling:
One-step SeedVR/Video upscalers in Comfy are VRAM-hungry; many stick to Topaz Video AI (local “patched” exists) or external tools, then re-encode. Online services exist too.
24GB VRAM helps; 12–16GB often need aggressive blockswap/offload, or upscaling in stages.

>>106115785
what about 3070

>>106115811
3.7/5

>Resolution/aspect and stability

WAN 480p models prefer one dimension at/near 480 (e.g., 832x480, 496x672). 512x512 square works OK. Pushing far off native 480 baseline degrades quality/stability.
720p models handle higher inputs better. 480p model can sometimes look better at near-720-ish if tuned, but generally move to 720p model when crossing ~640x640 effective workload.
Keep sides multiples of 64/128 for some attention optimizations (e.g., radial attention). Radial attention only works with Sage v1 currently; quality hit reported.

>Workflows, errors, installation gotchas

For ComfyUI:
Portable on Windows: install deps in python_embeded. Restart after installs/updates in Manager.
Place models in models/diffusion_models, LoRAs in models/loras, VAE where needed. Use correct loader nodes (native WAN nodes or Kijai’s wrappers).
Common errors:
“PassManager::run failed” after adding Sage/Triton: mismatched versions or wrong Python env. Use helper bundles or pinned versions.
“Reconnecting at Load Diffusion Model 62%”: out of RAM. Increase Windows pagefile (swap) to 64–200GB, but best to get 32–64GB RAM.
SageAttention not detected: wrong env vs Desktop vs portable mismatch. Use helper installer or ensure pip installs into the Comfy env (not system Python).
Slow lora/model reloads: reduce dynamic changes mid-run; prefer native workflows; Kijai published nodes to avoid full reloads when swapping LoRAs; avoid bf16 “fast” toggles that force recompile.

>>106115708
4/5
>Models, LoRAs, samplers

Lightx2V distill LoRAs:
New rank64 for T2V/I2V (official lightx2v repos) are strong. Rank64 is often optimal; rank128+ can hurt motion.
Kijai variants exist (rank64/128/256; adaptive quantile). Community reports: I2V rank64 is a “win”: more movement, less oily motion, more prompt sensitivity.
Some T2V Lightx2V rank64 had noise due to a bad release; fixes rolled, Kijai published alternatives.
Use matching LoRA type: T2V LoRA on I2V often causes slow-mo; vice versa too.
Strength: Many use 0.7–0.8 at 5 steps to avoid overcooking. 4 steps is very fast but can look artificial.
Other speed LoRAs:
PusaV1 (RaphaelLiu) + Kijai wrappers: mixed early impressions; feels different, sometimes weaker adherence; faster for T2V.
NAG:
Best used with native NAG ksampler (ComfyUI-NAG). Regular negative is ignored at cfg=1; feed “NAG-negative” properly. Gains are modest; can slow gen. Often better to weight positive tokens and skip complex negatives on cfg=1 Lightx2V runs.
Samplers:
LCM popular for fast 4–5 step WAN. UNI3C discussed, but it’s camera-control context, not a magical adherence boost.
FP8 notes:
fp8_e4m3fn_fast sometimes degrades quality; try non-fast. For RTX 30xx, use fp8_e5m2 weights if needed.

>>106115708
4.3/5
EXLUSIVE /adt/ WAN 2.1 INFO

Hardware, performance, and settings

>Typical speeds (optimized):

RTX 4060 16GB: ~2 minutes for 5s 480p with Lightx2V, 4–5 steps, blockswap 12–20.
RTX 4090: Commonly 1 minute or less for 5s 480p with Triton + SageAttention + Lightx2V.
RTX 2060/2060S: 15–20 minutes for 5s without accelerators; Lightx2V is essential to make it reasonable.
RTX 4070 Ti (12GB): Usable, but VRAM-limited for high res; still fast with Lightx2V.
RTX 3090 (24GB): Good balance; can run 720p decently and supports heavier workflows.
H100/A100: Not discussed with concrete numbers for WAN; overkill but fast in cloud.

>VRAM vs. chip:
Chip speed matters a lot. VRAM matters for resolution/context length/upscaling.
GGUF offload to RAM costs roughly 2× speed slowdown vs full VRAM, but still usable (not like LLMs).
32GB RAM is a minimum for comfort; 64GB is the sweet spot. Heavy workflows can blow past 32GB; swap helps but slows.
Blockswap vs. GGUF:
GGUF is easy but slower and can degrade quality vs safetensors. Blockswap lets you keep safetensors and only swap a few blocks to fit VRAM with minimal speed loss.
File swap/swap delays:
Large delays often come from reloading models/loras between runs. Kijai added nodes to reduce reloading, and native workflows help.
Linux vs Windows:
Linux can be finicky with fan curves and performance; manual scripts help. Windows is more plug-and-play.
CRF and video output:
Lower CRF = higher quality and bigger file. CRF=0 gives maximum quality (huge files). CRF=19 is a balanced default.

>>106115819
4/5
>>106115816
wait
>Comparisons with other video models

LTXV: fast/light but poor control and anatomy; produces lots of “random chaos.” Not recommended over WAN for quality/control.
Hunyuan-Video: heavier VRAM needs (old guidance ~45GB “max quality”); community moved on because WAN + Lightx2V is both faster and better for local use.
Kling 2.x (cloud): produces nice results; censorship/filters apply; not local and often paid.
Paid services: can be faster and high-quality out of the box, but filter heavy (NSFW blocks) and expensive for iteration. Many users prefer local for NSFW and full control.

>Content and censorship

Local = full freedom, but you must learn tools and optimize. Services = quality/speed, but strict content filters. Some services remove generations without refunds (e.g., VIDU Q1 anecdotes).
Safety: avoid sharing real people NSFW publicly; various jurisdictions cracking down on platforms (e.g., Civitai region blocks mentioned).

>>106115820

5/5
>Practical tips and FAQs

FPS and interpolation:
WAN natively trained around 16fps; if you need 30/60fps, use interpolation 2×/4× and set the Video Combine fps accordingly to avoid slow/fast playback.
Looping:
“First-Last Frame” i2v loops can produce artifacts (flash) near the loop point; better to inpaint/align transitions, or splice with VACE, or use FLF2V where available.
RAM/swap:
16GB RAM is too little for comfort. 32GB minimum; 64GB recommended. Swap works but slows everything; expect delays loading models/LoRAs.
MagRef resolution:
Works at 480p base; 512–656 squares are okay. 720p possible with more VRAM and careful settings.
Negative prompts with cfg=1:
With Lightx2V cfg distill and cfg=1, standard negative has minimal effect; you can zero it out. Use NAG’s own negative only if using NAG sampler.
Face/identity:
i2v MagRef is best for preventing “face leaking.” You can also add face ref/clip vision in some workflows or do a final faceswap (inswapper_128).
Slow-mo causes:
Mismatched LoRAs (T2V on I2V), LoRAs trained on slow footage, or mixing speed-up LoRAs incorrectly. Fix by using correct pairs and/or adding a “fast-motion” LoRA to counter.
Quality drift vs first frame:
Often due to incompatible LoRAs (e.g., T2V LoRA on I2V), low base resolution, or too low steps. Try 5–10 steps, use I2V rank64, or move to 720p model.
20xx GPUs and Sage:
Sage v2 limited; try older Sage, or skip. Radial attention only compatible with Sage v1 and specific size constraints.

>>106115816
>3070 its over
32–64GB system RAM is highly recommended; 12–24GB VRAM is comfortable.
For local T2V/I2V today, WAN 2.1 + Lightx2V + Triton + SageAttention + blockswap is the best combo. 4–5 steps are fine for speed; 10 steps give more adherence.
Use I2V MagRef for identity-stable edits; use T2V with Pusa/Lightx2V for faster motion.
Interpolate to 30/60fps after generation rather than hoping for native 30fps training.

tl;dr

>>106117186
Stop necrobumping faggot

>>106117218
Go back to whatever forum you came from, idiot.

>>106117223
Only if you stop being a loser

>>106117254
I'm more than happy to keep correctly labeling you what you are: a moron.

>>106117265
Well at least i'm not a snlowflake like you that is scared of actual women

>>106117286
Where do you think you are, idiot?

>>106117291
In a thread full of people that are scared to see realistic women?

>>106117298
It's good to see you're still upset over the mere existence of this splinter thread. Seeing a moron like you seethe is exactly why it should be kept alive.

>>106117304
Yeah i'm even bumping it once it's in danger to fall off the catalogue to keep your slop contained kek

Hi, what's character select saa?

>>106117403
USELESS THREAD ANSWER ME ITS BEEN AN HOUR YOU USELESS SPAM THREAD NO PURPOSE CANT EVEN ASNWER A QUESTION USELESS USELESS ULESUSS!!!
DELETE DELTE DELETE!!!

>>106118060
meds

Saar please tell me how to generate the anime woman with the bobs and vagene saar

>>106118103
nsfw, 1girl, nipples, gigantic breasts, pussy

Any good pooping gens?

what's the best model for inpainting anime?

Which anime model is the best for cute scat?