Velvet Chroma v2.0
Recommended settings:
Sampler: Euler / Deis2m
Scheduler: Beta / Beta57 / Sigmoid_offset
Steps: 10
CFG: 4
Hires Fix / Adetailer: recommended
Resolution: 1024x1024 / 832x1216 / 896x1152
Clip: Flant5-xxl_Q8_0 works fine
VAE: ae.safetensors
Recommended Prompting:
Trigger words: none
Negative prompt: none (or only worst quality, low quality)
Comfyui worflow: On showcase pictures or here
Velvet Chroma v1.0
Recommended settings:
Sampler: Euler · Steps: around 30 · CFG: 3–4
Hires Fix: recommended
Resolution: 896×1152 · VAE: ae.safetensors
Other quantizations can be added on request.
Description
v1.1 FP16
FAQ
Comments (9)
I have an RTX 4090. I checked the text encoders and there are some that are quite large, like 20GB or 45GB. Will using these improve the quality? Or do you have any recommendations?
Tthe larger Text Encoders (20GB–45GB) are generally the highest-quality ones. However, with a 24GB GPU like the RTX 4090 (i'm using a 5090, it's practically the same), you’re not really going to see any noticeable improvement in output quality, because the gain is minimal compared to the extra VRAM load. The best balance is to use a top-tier TextEncoder quantized (Q8 or Q6) or even just go straight to a FP16 build if available. It’ll save memory and give you virtually the same visual fidelity.
The main issue with Chroma right now is that image generation is slower compared to Flux, so every bit of VRAM you can save helps reduce generation time. Sometimes the time you save is worth more than the tiny quality loss. For example, going from FP16 to FP16_Q8_0 can reduce generation time from around 120 seconds to just 58 seconds at 832×1216 with Face Detailer and 4× upscale — a huge difference for almost no visible drop in quality.
For a 4090, I’d recommend this model:
flan-t5-xxl-fp16.safetensors: silveroxides/flan-t5-xxl-encoder-only
or a quantized version, Q8 or Q6 build: dumb-dev/flan-t5-xxl-gguf
@DeViLDoNia Thank you for your kind reply. It was a great help.
@DeViLDoNia i uses the chroma1 workflow template that comfyui already provided and uses text encoder fp16 as recommended. Wow the default setting at 1024x1024 with this velvet chroma already consume over 30gb vram. Luckily like u, i also have rtx 5090. mine is asus astral model. can i use text prompt similar to i have been using for illustrious in chroma1hd?
@yuto303 Yes, you definitely can. Chroma1 has a very strong understanding of Danbooru-style tags, and Velvet Chroma expands that knowledge even further, so your Illustrious-style prompts will work perfectly fine. With a 5090 you won’t have any trouble at all.
If you want to reduce VRAM usage, I recommend using the checkpoint quantized in Q8_0 along with the flan-t5-xxl Q8_0 text encoder. With that combination you’ll be able to generate images with Chroma while still using your PC normally at the same time.
@DeViLDoNia thanks. What about scheduler setting? Did you use this sigmoid scheduler? https://github.com/silveroxides/ComfyUI_SigmoidOffsetScheduler
I heard it is best for chroma.
Currently im using euler sampler and beta scheduler. Later tonight i will try to install this sigmoid scheduler and give it a try
@yuto303 I didn’t know about that scheduler either, but I’m installing it right now to test it as well.
@DeViLDoNia nice. Do let me know the result
@yuto303 it works well. I need to test it more, but i think it's probably a good one.