Search Results
7/12/2025, 9:43:11 AM
>>105878594
>This is why Chroma is so good, since it hasn't been trained 512x512.
You mean because it has been trained at 512x512, right ?
And yes it has, while the model I used for inference was the detail calibrated which is a merge with the 1024 res fork IIRC, the model I trained against was chroma-unlocked-v43.safetensors which I believe is 100% 512 trained (the last two epochs are planned to be bumped to 1024).
The 1024 fork has only been going for a couple of epochs, the 512 main branch has been going for 43 epochs.
That said, I'm unconvinced 1024 training is somehow responsible for fake/plastic skin, I think it all depends on the training images. Flux dev was trained on synthetic output from Flux Pro, that means it will look less realistic. Still, I am impressed at how much detail Chroma retains from training at mere 512 resolution, this speeds testing up SO MUCH.
>This is why Chroma is so good, since it hasn't been trained 512x512.
You mean because it has been trained at 512x512, right ?
And yes it has, while the model I used for inference was the detail calibrated which is a merge with the 1024 res fork IIRC, the model I trained against was chroma-unlocked-v43.safetensors which I believe is 100% 512 trained (the last two epochs are planned to be bumped to 1024).
The 1024 fork has only been going for a couple of epochs, the 512 main branch has been going for 43 epochs.
That said, I'm unconvinced 1024 training is somehow responsible for fake/plastic skin, I think it all depends on the training images. Flux dev was trained on synthetic output from Flux Pro, that means it will look less realistic. Still, I am impressed at how much detail Chroma retains from training at mere 512 resolution, this speeds testing up SO MUCH.
Page 1