<h2 id="what-this-is">What this is?</h2>My H3 workflow, designed to have quick generations on my relatively low VRAM (12gb 3060).On my setup, this workflow has the following generation times (compared to a basic 4 step workflow), rounded to the nearest 30 seconds:<pre><code>5 seconds, I2V
| Megapixels: | Basic 4 Step | V1.0 | V2.0 No Refine | V2.0 With Refine |
---------------------------------------------------------------------------
| 0.5 | 3:30 | 2:30 | 3:00 | 3:30 |
| 1.0 | 9:30 | 5:30 | 4:00 | 5:00 |
| 2.0 | 30:00 | 13:30 | 10:00 | 13:30 |

10 seconds, I2V
| Megapixels: | Basic 4 Step | V1.0 | V2.0 No Refine | V2.0 With Refine |
---------------------------------------------------------------------------
| 0.5 | 9:00 | 4:30 | 4:30 | 5:30 |
| 1.0 | 27:30 | 10:00 | 10:00 | 14:30 |
| 2.0 | OOM, DNF | 20:30 | 19:30 | 29:00 |</code></pre><h3 id="update-2.0:">Update 2.0:</h3>- Changed how the reference scaling works, now you can choose to match the aspect ratio of the generated video (for first/last frame/video editing) and separately scale the reference (for saving vram and increasing generation times).- Added the ability to trim video and audio references.- Video and audio references are now checked for length, an longer than 15s are trimmed to 15s, and if the total length exceeds 15s the references that push it over are disabled.- Added live generation preview. It always works for the hi-res steps, and sometimes works for the lo-res steps. I think it has to do with the final size whether it will preview the lo-res steps or not.- Added the option to do an extra, slower, refining step for better quality around moving things at the cost of generation time.<h2 id="features">Features</h2><h3 id="supports-both-ref2va-and-fl2va">Supports both ref2va and fl2va</h3>Select a fl2va model and toggle the 'use fl2va?' option to run in fl2va mode, leave it off and load a ref2va model for ref2va mode.<h3 id="automatically-constructs-prompts">Automatically constructs prompts</h3>Each section of a H3 prompt (subject_descriptions, retention_analysis, etc.) has it's own input, all of which are combined to construct the prompt, making it easier to keep track of everything.<h3 id="dynamic-prompts">Dynamic prompts</h3>Uses the <a target="_blank" rel="ugc" href="https://github.com/exectails/comfyui-et_dynamicprompts">dynamic prompts extension</a> to do some pretty neat things, the most useful in this case is using variables in your prompt.<h3 id="lora-support">LoRA Support</h3>Supports any LoRAs you want. Note: Adding LoRAs will increase generation time.<h3 id="easy-references">Easy references</h3>Supports the maximum amount of references ref2va uses (9 image, 3 video and 3 audio), and will automatically disable image references 9, 8 and 7 for each reference over the maximum of 12. Additionally, any video or audio references longer than 15s will be trimmed to 15s, and if the total length of all audio or video references exceeds 15s (Calculated separately for both video and audio) it will disable the ones that push it over.<h3 id="reference-control">Reference Control</h3>Allows the cropping of image and video references to match the target aspect ratio, additionally, allows image and video references to be scaled to save vram and decrease generation time (Especially useful when doing vid 2 vid stuff). Video and audio references can also be trimmed to only use the section most relevant to the generation.<h3 id="refine-step">Refine step</h3>An extra, slower refining step can be enabled to clean up motion around fast moving objects, especially noticeable with things like hair. It is not always needed, but it does make a difference.<h2 id="required-node-packs">Required node packs</h2>Uses the following node packs:<pre><code>🟢 Will be auto-detected by comfyui-manager
🟡 Can be installed using comfyui-manager, but is not auto-detected (you will have to search for it). Manual installation may be easier to avoid getting the wrong pack.
🔴 Requires manual installation</code></pre>🟢 <a target="_blank" rel="ugc" href="https://github.com/ltdrdata/ComfyUI-Impact-Pack">Impact Pack</a>🟢 <a target="_blank" rel="ugc" href="https://github.com/kijai/ComfyUI-KJNodes">KJ Nodes</a>🟢 <a target="_blank" rel="ugc" href="https://github.com/rgthree/rgthree-comfy">RGThree Comfy</a>🟢 <a target="_blank" rel="ugc" href="https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite">Video Helper Suite</a>🟢 <a target="_blank" rel="ugc" href="https://github.com/dailydoseofaiart/ComfyUI-FPSFrameDrop">FPS Frame Drop</a>🟡 <a target="_blank" rel="ugc" href="https://github.com/exectails/comfyui-et_dynamicprompts">Dynamic Prompts</a> Note: If installing through comfyui-manager, get the one with the description "Nodes that implement functionality similar to the Dynamic Prompts extension for A1111"🟡 <a target="_blank" rel="ugc" href="https://github.com/Fictiverse/ComfyUI_Fictiverse">Fictiverse</a> Nodes Note: The nodes from this pack are incorrectly detected as belonging to another pack. Make sure you have this one installed before installing auto detected packs.🔴 <a target="_blank" rel="ugc" href="https://github.com/T8mars/comfyui-minimax-h3-audio-T8">MiniMax H3 Audio T8</a> Note: This one can be installed using comfyui-manager (it is not auto-detected), but reports it's compatibility requirements incorrectly so it cannot easily be enabled. Manual installation bypasses that.🔴 <a target="_blank" rel="ugc" href="https://github.com/yichengup/ComfyUI-YCNodes-MiniMax-H3">YC Nodes MiniMax H3</a>🔴 <a target="_blank" rel="ugc" href="https://github.com/slmonker/selflift-Avatar">Selflift Avatar</a> Important note: This packs latest update has broken it somehow, either use <code>git checkout 8e01c09</code> if you've already download it to revert to an older commit where it still works correctly, or if you're downloading for the first time use <code>git clone --revision=8e01c093f027c0281025741b025f117475ea044b https://github.com/slmonker/selflift-Avatar.git</code> when installing. Note 2: This workflow was made with an older version of the pack, either revert to commit <code>dc8e545</code> to use the workflow as is, or recreate the 'selflift-avatar sampler' node, toggle the 'hires_tiling' option and select 'spatial tiling (legacy)' for 'hires_tiling_mode' if you want to use a more up to date version.<h2 id="required-models">Required models</h2><h3 id="diffusion-model">Diffusion model</h3>Any int8-convrot model, I use <a target="_blank" rel="ugc" href="https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models">these hybrid models</a> or <a target="_blank" rel="ugc" href="https://civarchive.com/models/2900456/minimax-h3sparseref15hybrid?modelVersionId=3301688">this one</a> for ref2va, and the <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors">basic fl2va one</a> or <a target="_blank" rel="ugc" href="https://civarchive.com/models/2877206/dasiwa-minimax-h3">this one</a> for fl2va. These go in models/diffusion_models<h3 id="text-encoder">Text encoder</h3><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors">This one</a>. Goes in models/text_encoders<h3 id="vae">Vae</h3><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_video_vae_int8_convrot.safetensors">Video</a>, <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_audio_vae_fp32.safetensors">audio</a>. Goes in models/vae<h3 id="latent-upscaler">Latent upscaler</h3><a target="_blank" rel="ugc" href="https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler/blob/main/minimax_h3_latent_upscaler_3d_conv_v1/minimax_h3_latent_upscaler_3d_conv_v1_bf16.safetensors">This one</a>. Goes in models/latent_upscale_models<h3 id="4-step-lora">4 Step LoRA</h3>Any pruned LoRA model (must be pruned, will crash otherwise). I use <a target="_blank" rel="ugc" href="https://huggingface.co/Kijai/MiniMax-H3-experimental/blob/main/loras/minimax_h3_4step_lora_flashgen_v1.0_768p_fl2va_pruned_avg_rank_13_bf16.safetensors">this one</a>, <a target="_blank" rel="ugc" href="https://civarchive.com/models/2837571/minimax-h3-turbo-loras?modelVersionId=3202732">this one</a> is also pretty good. Goes in models/loras<h3 id="full-model-paths">Full model paths</h3><pre><code>📂 ComfyUI
├── 📂 models
| ├── 📂 diffusion_models
| | ├── DIFFUSION MODELS GO HERE
| |
| ├── 📂 latent_upscale_models
| | ├── LATENT UPSCALER GOES HERE
| |
| ├── 📂 loras
| | ├── 4 STEP LORA GOES HERE
| |
| ├── 📂 text_encoders
| | ├── TEXT ENCODER GOES HERE
| |
| ├── 📂 vae
| | ├── VAE MODELS GO HERE</code></pre>

MiniMax H3 ref2va/fl2va fast generation workflow

Update 2.0:- Changed how the reference scaling works, now you can choose to match the aspect ratio of the generated video (for first/last frame/video editing) and separately scale the reference (for saving vram and increasing generation times).- Added the ability to trim video and audio references.- Video and audio references are now checked for length, an longer than 15s are trimmed to 15s, and if the total length exceeds 15s the references that push it over are disabled.- Added live generation preview. It always works for the hi-res steps, and sometimes works for the lo-res steps. I think it has to do with the final size whether it will preview the lo-res steps or not.- Added the option to do an extra, slower, refining step for better quality around moving things at the cost of generation time.FeaturesSupports both ref2va and fl2vaSelect a fl2va model and toggle the 'use fl2va?' option to run in fl2va mode, leave it off and load a ref2va model for ref2va mode.Automatically constructs promptsEach section of a H3 prompt (subject_descriptions, retention_analysis, etc.) has it's own input, all of which are combined to construct the prompt, making it easier to keep track of everything.Dynamic promptsUses the <a target="_blank" rel="ugc" href="https://github.com/exectails/comfyui-et_dynamicprompts">dynamic prompts extension</a> to do some pretty neat things, the most useful in this case is using variables in your prompt.LoRA SupportSupports any LoRAs you want. Note: Adding LoRAs will increase generation time.Easy referencesSupports the maximum amount of references ref2va uses (9 image, 3 video and 3 audio), and will automatically disable image references 9, 8 and 7 for each reference over the maximum of 12. Additionally, any video or audio references longer than 15s will be trimmed to 15s, and if the total length of all audio or video references exceeds 15s (Calculated separately for both video and audio) it will disable the ones that push it over.Reference ControlAllows the cropping of image and video references to match the target aspect ratio, additionally, allows image and video references to be scaled to save vram and decrease generation time (Especially useful when doing vid 2 vid stuff). Video and audio references can also be trimmed to only use the section most relevant to the generation.Refine stepAn extra, slower refining step can be enabled to clean up motion around fast moving objects, especially noticeable with things like hair. It is not always needed, but it does make a difference.Required node packsUses the following node packs:<pre><code>🟢 Will be auto-detected by comfyui-manager
🟡 Can be installed using comfyui-manager, but is not auto-detected (you will have to search for it). Manual installation may be easier to avoid getting the wrong pack.
🔴 Requires manual installation</code></pre>🟢 <a target="_blank" rel="ugc" href="https://github.com/ltdrdata/ComfyUI-Impact-Pack">Impact Pack</a>🟢 <a target="_blank" rel="ugc" href="https://github.com/kijai/ComfyUI-KJNodes">KJ Nodes</a>🟢 <a target="_blank" rel="ugc" href="https://github.com/rgthree/rgthree-comfy">RGThree Comfy</a> 🟢 <a target="_blank" rel="ugc" href="https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite">Video Helper Suite</a>🟢 <a target="_blank" rel="ugc" href="https://github.com/dailydoseofaiart/ComfyUI-FPSFrameDrop">FPS Frame Drop</a>🟡 <a target="_blank" rel="ugc" href="https://github.com/exectails/comfyui-et_dynamicprompts">Dynamic Prompts</a> Note: If installing through comfyui-manager, get the one with the description "Nodes that implement functionality similar to the Dynamic Prompts extension for A1111"🟡 <a target="_blank" rel="ugc" href="https://github.com/Fictiverse/ComfyUI_Fictiverse">Fictiverse</a> Nodes Note: The nodes from this pack are incorrectly detected as belonging to another pack. Make sure you have this one installed before installing auto detected packs.🔴 <a target="_blank" rel="ugc" href="https://github.com/T8mars/comfyui-minimax-h3-audio-T8">MiniMax H3 Audio T8</a> Note: This one can be installed using comfyui-manager (it is not auto-detected), but reports it's compatibility requirements incorrectly so it cannot easily be enabled. Manual installation bypasses that.🔴 <a target="_blank" rel="ugc" href="https://github.com/yichengup/ComfyUI-YCNodes-MiniMax-H3">YC Nodes MiniMax H3</a>🔴 <a target="_blank" rel="ugc" href="https://github.com/slmonker/selflift-Avatar">Selflift Avatar</a>Required modelsDiffusion modelAny int8-convrot model, I use <a target="_blank" rel="ugc" href="https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models">these hybrid models</a> for ref2va, and the <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors">basic fl2va one</a>. These go in models/diffusion_modelsText encoder<a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors">This one</a>. Goes in models/text_encodersVae<a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_video_vae_int8_convrot.safetensors">Video</a>, <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_audio_vae_fp32.safetensors">audio</a>. Goes in models/vaeLatent upscaler<a target="_blank" rel="ugc" href="https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler/blob/main/minimax_h3_latent_upscaler_3d_conv_v1/minimax_h3_latent_upscaler_3d_conv_v1_bf16.safetensors">This one</a>. Goes in models/latent_upscale_models4 Step LoRAAny pruned LoRA model (must be pruned, will crash otherwise). I use <a target="_blank" rel="ugc" href="https://huggingface.co/Kijai/MiniMax-H3-experimental/blob/main/loras/minimax_h3_4step_lora_flashgen_v1.0_768p_fl2va_pruned_avg_rank_13_bf16.safetensors">this one</a>, <a target="_blank" rel="ugc" href="https://civitai.red/models/2837571/minimax-h3-turbo-loras?modelVersionId=3202732">this one</a> is also pretty good. Goes in models/lorasFull model paths<pre><code>📂 ComfyUI
├── 📂 models
| ├── 📂 diffusion_models
| | ├── DIFFUSION MODELS GO HERE
| |
| ├── 📂 latent_upscale_models
| | ├── LATENT UPSCALER GOES HERE
| |
| ├── 📂 loras
| | ├── 4 STEP LORA GOES HERE
| |
| ├── 📂 text_encoders
| | ├── TEXT ENCODER GOES HERE
| |
| ├── 📂 vae
| | ├── VAE MODELS GO HERE</code></pre>

Files with SHA256: 44f09673908bf194fd5c296adff01cde5e9e42e2d261b7ccf3c85de9c3f0afe9