A well trained character lora greatly improves the output. All of my examples are without a character Lora.
If you dont notice a difference from nag scale and nag tau, you need to lower nag sigma end.
## MODELS - For 8GB VRAM or more
[Chroma1-HD-Q8_0](https://huggingface.co/silveroxides/Chroma1-HD-GGUF/blob/main/Chroma1-HD-Q8_0.gguf)
or Q6_k for 6gb cards [Chroma1-HD-Q6_K](https://huggingface.co/silveroxides/Chroma1-HD-GGUF/blob/main/Chroma1-HD-Q6_K.gguf)
## Text_encoder
[FLAN-T5-XXL Q8 GGUF](https://huggingface.co/easygoing0114/flan-t5-xxl-fused/blob/main/flan_t5_xxl_TE-only_Q8_0.gguf)
## VAE - this version saves vram
[flux-vae-bf16](https://huggingface.co/Kijai/flux-fp8/blob/main/flux-vae-bf16.safetensors)
## LORA
if you don't like the quality or want less steps pick a higher rank
Standard flash
[chroma-flash-lora_r16-fp32.safetensors](https://huggingface.co/silveroxides/Chroma-LoRAs/blob/main/chroma-flash/chroma-flash-lora_r16-fp32.safetensors)
Seems to do better realism but needs more steps
[Chroma1-Flash_r12-fp32.safetensors](https://huggingface.co/silveroxides/Chroma-LoRAs/blob/main/Chroma1-Flash/Chroma1-Flash_r12-fp32.safetensors)
Optional realism lora
[lenovo-ultrareal](https://civarchive.com/models/1662740/lenovo-ultrareal)
## NOTES
Changed to targeting DPMPP_2M + Simple, This allows for quicker convergence and higher resolution and virtually no blocking.
Adding a well trained Character lora significantly improves the quality of outputs
720x1280 now targeted resolution, smaller resolutions like 640x960 work great as well.
## EXTRA prompt notes
Realistic
_________________
high resolution, professional photograph,
soft cinematic lighting, warm tones, soft bokeh.
_________________
Amateur
_________________
DCIM, webcam, amateur photograph, film grain, accidental shot, candid photography, low saturation,
overall candid and relaxed aesthetic, low contrast, neutral lighting, low saturation, soft bokeh, flat composition, calm tones. Shot on an iPhone. The image shows signs of compression.
Description
This version removes the use of RES4LYF as it has started causing memory leaks on some systems.
Changed to targeting DPMPP_2M + Simple, This allows for quicker convergence and higher resolution and virtually no blocking.
Adding a well trained Character lora significantly improves the quality of outputs
720x1280 now targeted resolution, smaller resolutions like 640x960 work great as well.
technically 6GB cards can now run the Q8 model
FAQ
Comments (5)
Thanks for sharing your workflow!
I had one issue trying to get it to work where the 'bong_tangent' scheduler wasn't found, and ComfyUI-Manager wasn't able to determine which node was missing. Installing the custom node pack 'RES4LYF' manually fixed it.
Hello! I tried the models in Forge, they work great, thank you, I would never have found them myself.
At the same time, I still wanted to work with ComfyUI, because it feels a little better optimized, but I ran into a problem where I can't download the missing nodes using the manager. If it's not too much trouble, could you specify which packs I need?
I'm getting error NAGCFGGuider.inner_sample() got an unexpected keyword argument 'latent_shapes', anyone got the same?
Thanks for the workflow. This became my main chroma wf after adding a few things like face detailer and post process
NAG module won't install
[DEPRECATION WARNING] Detected import of deprecated legacy API: /scripts/ui/components/button.js. This is likely caused by a custom node extension using outdated APIs. Please update your extensions or contact the extension author for an updated version.
Looks like we don't have an active mirror for this file right now.
CivArchive is a community-maintained index — we catalog mirrors that volunteers upload to HuggingFace, torrents, and other public hosts. Looks like no one has uploaded a copy of this file yet.
Some files do get recovered over time through contributions. If you're looking for this one, feel free to ask in Discord, or help preserve it if you have a copy.