@iamddtla thx for your comment, do you have any tutorial do convert it?

@DarkShiva gguf models nowadays are not worth it if you plan to use them in ComfyUI. They are not compatible with the new memory management system, can create many issues. Safetensors are the best choice in general, even when the models are heavier than ggufs in term of size.
Currently the best choice for file size, quality and speed are int8 models. They basically keeps the original quality of the full model, and are very fast in generation. Very recommended.

@DarkShiva I can share with you the V1 Q4_0 GGUF (converted from FP8). It's not the best choice, but it's usable.

@LG89 It isn't necessarily about that is it? On CUDA with plenty of VRAM, BF16 may be the best choice. On unified-memory systems or other environments where memory pressure is the limiting factor, GGUF may be the more practical deployment. Try removing your blinkers.

@mrmrswiggly612 I said that because many people are experiencing issues with GGUFs in ComfyUI. Comfyanonymous himself recommends to stop using that format because it's not working well with dynamic vram and all the new memory optimizations they made. This is why I recommended to use safetensors instead, and int8 is currently the most efficient quantisation in my opinion. Obviously I can only speak for Nvidia low-vram users (8GB), GGUFs never worked good for me since the beginning and I kept using fp16, bf16 or mxfp8 quants in most cases thanks to the offloading.

You don’t need to keep arguing here. For me, GGUFs are the best option I can use at home in terms of speed and quality. I even tested it again yesterday, comparing the original Turbo INT8 version with the original Turbo Q8 GGUF, and there is no difference in quality. The only difference is that the INT8 version takes about ten times longer to process — instead of getting 16 seconds per iteration, I get 211 seconds per iteration.

For me, that’s simply not an option, but everyone should decide for themselves. And if the author doesn’t want to provide GGUF versions, then that’s their choice — it’s not the end of the world. There are plenty of checkpoints available in GGUF format anyway.

@DarkShiva If ever you find the GGUF version you're searching for or find a way to convert/ quantize it, thanks to let me know! It's also more practical in my case. Thanks!

@Bidi_Skimmer no, but i dont care, i use the raw with snofs that give me good results. better than any turbo

@iamddtla 有用lora,我是直接用的之前fp8照片的数据,现在更多会出现很多噪点,这让我很头疼

@iamddtla 我不太清楚,但是这个现象偶然发生,我也不想去管了

我也是

Same. I genuinely hope the author considers uploading the model on Huggingface or any other platform after the early access is over.

已经可以下载了

me 2

已经可以下载了

yes, use regular krea2 workflow, chain 2 ksamplers first to unsample second to resample, changing denoise level you can adjust how far it variates from original