>>105962792

>CHROMA YUME NOOBAI XL

• What it is
A V-prediction diffusion model (NOOBAI XL-VPred lineage) rebuilt across 4 numbered releases. Only runs on tensor.art for now (not Civitai).

• Data & training highlights
– v1.0–3.0: Danbooru 2024, Yande, e621 + Illustrious XL & NOOBAI XL as teachers
– v2.0: +50 k real-life photos
– v3.0: dataset relabeled with GPT-4.5 then manually checked
– v4.0: fresh train on expanded anime/photographic sets (danbooru_newest, gelbooru_full, etc.) + custom captioned data; CLIP, VAE and U-Net retrained from scratch

• Version-to-version behavior
v1.0 – balanced multi-style anime, solid anatomy
v2.0 – much better anatomy & realism, weaker cross-style quality
v3.0 – recovers multi-style capability while keeping realism; prompt accuracy now critical
v4.0 – cleaner anatomy and style faithfulness, resolves v2/3 issues

• Core usage tips (V-Pred specific)
Positive (generic): “masterpiece, best quality, amazing quality”
Negative (generic): “bad quality, worst detail, sketch, censor, simple/transparent background”
Typical hyper-params:
‑ Sampler: Euler a (or Euler Ancestral CFG++ –Simple)
‑ Steps: 20 – 30 (25 – 30 with CFG++ recipe)
‑ CFG: 4 – 6 (1.2 – 1.5 with CFG++ recipe)
‑ Clip-skip: 2

For photoreal / cosplay output add to prompts:
Positive: “realistic, cosplay, real life, photorealistic”
Negative: “illustration, blur, film grain, cartoon, anime coloring, 3D, 2D, unrealistic …”

• Misc.
– All showcase images are straight from the base model (no LoRA / post-processing).
– Author actively seeks feedback to refine future versions.
– Credits: narugo1992, Nyanko, Laxhar Lab, Sennke and others.

https://civitai.com/models/1330192/chromayume-noobai-xl-nai-xl