Search Results
7/24/2025, 9:03:09 PM
>JoyCaption Beta One
>CLIP-KO
>VAE EQ (MS‑LC‑EQ‑D‑VR VAEFLUX)
The new stuff used in my last training session.
JoyCaption is still mildly retarded, but I had to fix less of them this time, CLIP-KO was used alongside the VAE EQ in training and it converged about 1200 steps earlier (1500 vs. 2700 with the defaults), but was also a lot slower* (about 75min per 1k steps on a 4090). There were also artifacts from the EQ process? (aliasing, oversharpening and a fine crosshatch pattern), though there's a possibility that this could also be from the bucketing process (it scales images extremely poorly).
*It was actually about the same speed, but I had no idea at the time it had converged that early and wound up going another ~1500 steps beyond what was needed.
Unfortunately, the LoRA also no longer overpowers the slopped look of base Flux Dev using these options. Maybe dropping back down to 512 for training will reduce that crisp/smooth look now that it converges so much better.
>CLIP-KO
>VAE EQ (MS‑LC‑EQ‑D‑VR VAEFLUX)
The new stuff used in my last training session.
JoyCaption is still mildly retarded, but I had to fix less of them this time, CLIP-KO was used alongside the VAE EQ in training and it converged about 1200 steps earlier (1500 vs. 2700 with the defaults), but was also a lot slower* (about 75min per 1k steps on a 4090). There were also artifacts from the EQ process? (aliasing, oversharpening and a fine crosshatch pattern), though there's a possibility that this could also be from the bucketing process (it scales images extremely poorly).
*It was actually about the same speed, but I had no idea at the time it had converged that early and wound up going another ~1500 steps beyond what was needed.
Unfortunately, the LoRA also no longer overpowers the slopped look of base Flux Dev using these options. Maybe dropping back down to 512 for training will reduce that crisp/smooth look now that it converges so much better.
Page 1