Final version?
Is chroma-unlocked-v50.safetensors the final version, or should we expect chroma-unlocked-v50-detail-calibrated.safetensors, or even something after these two?
As of ~35 mins ago(at time of writing: Fri 8 Aug 11:25 BST 2025)
V49 and V50 is up!!!
(Asside: bit irregular its two at once, but not complaining)
https://huggingface.co/lodestones/Chroma/blob/main/chroma-unlocked-v49.safetensors
https://huggingface.co/lodestones/Chroma/blob/main/chroma-unlocked-v50.safetensors
Idk all i know is lodestones said "48-50" would be where the high-res training happens as before that its 512x512 training sets, it was mentioned in a discussion here about using Chroma for high-res stuff, but in the free time i had to search here in discussions, i couldn't find it, probably my error.
Maybe someone could link me to the discussion if they know.
yeah v50 is the final "base" model, speed model will come after
Sorry if I am asking a stupid question, but what is the difference between v50 and v50-annealed?
Yeah I wonder about that as well
Wouldn’t be better to name the v50 just Chroma since it’s done? Also will there be an official guide on how use the model at its best? Im thinking also all the custom nodes that exponentially increased its potential. So that newcomers who will start prompting “1girl, most beautiful alive, boobs” won’t start whining all over the internet of how bad the model is.
Sorry if I am asking a stupid question, but what is the difference between v50 and v50-annealed?
I don't have the exact answer but I can say that the term can have to do with DNA recombination so maybe some other sort of recombination is happening with this.
Has anything setting wise changed in V50, while v48 seemed fine now I get blurry pictures and what not.
后续是否能指导lora的训练，截至到现在还没看到相关的教程和案例？第二，是否可以添加controlnet支持？模型稍微可控应用会更多吧？
后续是否能指导lora的训练，截至到现在还没看到相关的教程和案例？第二，是否可以添加controlnet支持？模型稍微可控应用会更多吧？
A pose controlnet would also be amazing. I have a feeling we're going to be seeing more stuff like this for Chroma now that it's on v50.
后续是否能指导lora的训练，截至到现在还没看到相关的教程和案例？第二，是否可以添加controlnet支持？模型稍微可控应用会更多吧？
The controlnet have already been supported by ComfyAnonimous by more than a month or so and you can use this without any problems : https://huggingface.co/Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro-2.0.
As for the Lora someone already trained a couples as you would have made using Flux and they works great
Has anything setting wise changed in V50, while v48 seemed fine now I get blurry pictures and what not.
The Vae changed. Download the right one from https://huggingface.co/lodestones/Chroma1-HD
The VAE has not changed. It still uses Flux's VAE
what does annealed mean?
后续是否能指导lora的训练，截至到现在还没看到相关的教程和案例？第二，是否可以添加controlnet支持？模型稍微可控应用会更多吧？
The controlnet have already been supported by ComfyAnonimous by more than a month or so and you can use this without any problems : https://huggingface.co/Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro-2.0.
As for the Lora someone already trained a couples as you would have made using Flux and they works great
我很早就尝试了，不知道是设置还是参数问题，生成的都是噪点，类似牛仔面料放大的的那种噪点。一次都没成功过。有谁有可用的工作流吗？
what does annealed mean?
I think anneled version supports more diversity outputs. Higher temperature sort of thing, also like the DNA thing from previous comment mentioned.
后续是否能指导lora的训练，截至到现在还没看到相关的教程和案例？第二，是否可以添加controlnet支持？模型稍微可控应用会更多吧？
The controlnet have already been supported by ComfyAnonimous by more than a month or so and you can use this without any problems : https://huggingface.co/Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro-2.0.
As for the Lora someone already trained a couples as you would have made using Flux and they works great我很早就尝试了，不知道是设置还是参数问题，生成的都是噪点，类似牛仔面料放大的的那种噪点。一次都没成功过。有谁有可用的工作流吗？
I had the same issue when trying that one. Hoping someone can share a simple workflow for posing. I checked on the discord and didn't see any solutions there either.