<h1 id="hunyuan-video-coun7j5pd"><a target="_blank" rel="ugc" href="https://github.com/kijai/ComfyUI-HunyuanVideoWrapper">Hunyuan Video</a></h1><h3 id="kijai-marked-files-only-for-use-with-kijai-nodes-you-do-not-need-them-for-comfy-native-b9kn0q23m">Kijai marked files only for use with <a target="_blank" rel="ugc" href="https://github.com/kijai/ComfyUI-HunyuanVideoWrapper">Kijai Nodes </a>You do not need them for Comfy Native</h3><ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/articles/13522/hunyuan-video-what-model-to-use-for-8gb-24gb">Full Guide</a> to picking the correct file above</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/api/download/attachments/288942">Workflow </a>for 8GB Card users</li><li>Uncensored llama will work with COMFY Native</li></ul>Using the Kajai marked models on COMFY native will cause rainbow or black output.I do not recommend the the FP8 VAE unless you are trying to fit all models into GPU, see the guide for 4090 full GPU launch commands.<hr />Technical details regarding "Uncensored"The model used for Hunyuan was based on llava-llama-3 8 billion parameter LLM. The Intel vision tuned model was used to refine the tokenized model restoring over 5 million values.

Hunyuan Video (Safetensors) - New Uncensored Llama

hunyuanVideoSafetensors_visionCLIPLBF16.safetensors

Vision_CLIP-L-BF16

Uncensored_Llama_TE-BF16

Uncensored_Llama_TE-FP32

Scaled_Vision_CLIP-FP8

Comfy-VAE-BF16

CLIP-L_Large_FP8

Comfy-Diffusion-BF16

Comfy-VAE-FP32

Comfy-VAE-FP8

Comfy-Diffusion-FP8

Kijai-llava-llama-3-FP8

Kijai-llava-llama-3-BF16

Kijai-Diffusion-FP8

Kijai-Diffusion-BF16

Llava Llama TE

kijai-VAE-FP32

<h1 id="flux-blockwise-(mixed-precision-model)-x1l8hjix2">Flux Blockwise (Mixed Precision Model)</h1>I had to build several custom tools to allow for the mixed precision model, to my knowledge it is the first built like this.<ul><li>Faster and more accurate then any other FP8 quantized model currently available</li><li>Works in Comfy and Forge but forge needs to be set to BF16 UNET</li><li>Comfy load as a diffuser model USE DEFAULT WEIGHT</li><li>FP16 Upcasting should not be used unless absolutely necessary such as running CPU or IPEX</li><li>FORGE - set COMMANDLINE_ARGS= --unet-in-bf16 --vae-in-fp32</li><li>Other then the need to force forge into BF16, (FP32 VAE optionally) it should work the same as the DEV model with the added benefit of being 5GB smaller then the full BF16</li></ul>It turns out that every quantized model including my own up to this point to my knowledge has been built UN-optimally per blackforest.Only the UNET blocks should be quantized in the diffuser model, also they should be upcast to BF16 and not FP16 (Comfy does this correctly)<hr /><a target="_blank" rel="ugc" href="https://civarchive.com/user/White_Rabbit_A">Hippo Image remix</a><a target="_blank" rel="ugc" href="https://civarchive.com/user/Qvoheu">Lion Image remix</a>I am currently trying to workout how to follow Blackforest recommendations but using GGUF

Files with SHA256: 4a0bbab8efdee647bb35343a6ca48042546cefbbff315c1c6eb67cfde755fbe4