MiniMax H3 V2V Multi-Pass Pipeline — For Patreon Supporters (Patreon)
Videos
-
2026-08-11 04-47-56.mp4
Downloads
Content
Hey everyone. The public post has the free V2V workflow https://www.patreon.com/aifuturetech/posts/minimax-h3-video-166278462 — that one runs H3 Ref2V into LTX upsampler in a single chain.
This post has the advanced version. It adds SeedVR2 and a third sampling pass between H3 and LTX, then finishes with FILM frame interpolation and RTX Video Super Resolution for 4K output.
The reason this exists: MiniMax H3 can produce blurred or pixelated faces on mid and far shots. The single-pass H3-to-LTX chain helps, but edges on objects and characters are still not fully refined. Adding SeedVR2 as an intermediate upscaler sharpens those edges before LTX gets them. LTX then smooths the motion flickering that SeedVR2 can introduce. The final FILM plus RTX pass resolves the pixel density to genuine 4K.
----------------------------------------
WHAT'S INCLUDED
----------------------------------------
File attached:
1. MiniMax-H3 V2V Object Swap (Advance) -202608.json
1. MiniMax H3 V2V Multi-Pass Upscale Pipeline
.........................................
File: MiniMax-H3 V2V Object Swap (Advance) -202608.json
76 nodes. Four stages, each feeding the next:
Stage 1 — MiniMax H3 Ref2V with Turbo LoRA
- MiniMaxH3ReferenceToVideo — 1344x768, 5s, match mode
- Turbo LoRA: minimax-h3\minimax_h3_fl2v_lightx2v_turbo_4step_v0.1_comfy.safetensors at 0.7
- UNETLoader: minimax-h3\minimax_h3_ref2va_int8_convrot.safetensors
- CLIPLoader: minimax-h3\qwen3vl_32b_minimax_h3_int8_convrot.comfy.safetensors (type: minimax)
- Video VAE: minimax-h3\minimax_h3_video_vae_fp16.safetensors
- Audio VAE: minimax-h3\minimax_h3_audio_vae_fp32.safetensors
- MiniMaxH3SigmaShift: shift 12, sigma 6
- PathchSageAttentionKJ: auto
- Sampler: SamplerCustomAdvanced + KSamplerSelect (euler) + BasicScheduler (beta, 4 steps, denoise 1.0)
- ResolutionSelector: 9:16 Portrait Widescreen at 0.6 megapixels, multiple 32
- ControlNet preprocessors: DWPreprocessor (512) and DepthAnythingV2Preprocessor (512)
- Input: 2 reference images (transparent background character + environment), 1 reference video
- Image resize via ImageResizeKJv2 (1024x1024 for reference images, 512x512 for video frames)
Stage 2 — SeedVR2 Video Upscaler (edge sharpening pass)
- SeedVR2LoadDiTModel: seedvr2_ema_7b_fp16.safetensors (7B model)
- SeedVR2LoadVAEModel: ema_vae_fp16.safetensors
- SeedVR2VideoUpscaler: seed 42 fixed, target resolution 720, resample 5, batch processing
- Color space: lab
- This stage takes the raw H3 output and sharpens edges on faces, clothing boundaries, and objects. It runs between H3 and LTX so LTX gets a cleaner, higher-detail input to work with.
Stage 3 — LTX 2.3 Distilled Upsampler (second pass)
- CheckpointLoaderSimple: ltx2\ltx-2.3-22b-dev.safetensors
- Distilled LoRA: ltx2\ltx-2.3-22b-distilled-lora-384.safetensors at 1.0
- LTXAVTextEncoderLoader: gemma_3_12B_it_fp8_scaled.safetensors
- LTXVAudioVAELoader: ltx2\ltx-2.3-22b-dev.safetensors
- Video VAE: LTX23_video_vae_bf16.safetensors
- ImageScaleToMaxDimension: lanczos, 1920 max
- This stage smooths the motion flickering that SeedVR2 can introduce. The pixel detail from SeedVR2 is preserved, but temporal coherence is restored.
Stage 4 — Frame Interpolation + RTX 4K Upscale
- FrameInterpolationModelLoader: film_net_fp16.safetensors
- FrameInterpolate: multiplier 2 (doubles frame count)
- RTXVideoSuperResolution: scale by multiplier 2, quality ULTRA
- VHS_VideoCombine: 24fps output, h264-mp4, CRF 19
- Cache management: 3x easy cleanGpuUsed, 2x easy clearCacheAll between heavy stages
Pipeline flow:
Refs + prompt + video
-> MiniMaxH3ReferenceToVideo (Turbo LoRA 0.7 + SageAttention + SigmaShift 12/6 + DWProcessor + DepthAnythingV2)
-> SeedVR2 7B upscale (edge sharpening, lab color, target 720)
-> LTX 2.3 distilled second pass (flicker smoothing, max dim 1920)
-> FILM x2 (frame interpolation)
-> RTX VSR x2 ULTRA (4K resolution)
-> export 24fps h264
----------------------------------------
WHAT YOU NEED
----------------------------------------
MiniMax H3:
https://huggingface.co/Comfy-Org/MiniMax-H3
models/diffusion_models/minimax-h3/minimax_h3_ref2va_int8_convrot.safetensors
models/text_encoders/minimax-h3/qwen3vl_32b_minimax_h3_int8_convrot.comfy.safetensors
models/vae/minimax-h3/minimax_h3_video_vae_fp16.safetensors
models/vae/minimax-h3/minimax_h3_audio_vae_fp32.safetensors
Turbo LoRA:
models/loras/minimax-h3/minimax_h3_fl2v_lightx2v_turbo_4step_v0.1_comfy.safetensors
https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main
or
https://huggingface.co/Kijai/MiniMax-H3_comfy/tree/main/loras
SeedVR2:
https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler
models/checkpoints/seedvr2_ema_7b_fp16.safetensors
models/vae/ema_vae_fp16.safetensors
LTX 2.3:
models/checkpoints/ltx2/ltx-2.3-22b-dev.safetensors
models/loras/ltx2/ltx-2.3-22b-distilled-lora-384.safetensors
models/text_encoders/gemma_3_12B_it_fp8_scaled.safetensors
models/vae/LTX23_video_vae_bf16.safetensors
https://huggingface.co/Lightricks/LTX-2.3
Frame Interpolation + RTX:
https://huggingface.co/Comfy-Org/frame_interpolation
models/upscale_models/film_net_fp16.safetensors
https://github.com/Comfy-Org/Nvidia_RTX_Nodes_ComfyUI
NVIDIA RTX GPU required for RTX Video Super Resolution
GPU requirements:
16GB VRAM minimum to run the full chain. 24GB recommended for the SeedVR2 plus LTX stages without splitting.
On tighter cards, run stages separately with cache clearing between them.
----------------------------------------
QUICK START
----------------------------------------
1. Download all model files listed above before opening the workflow.
2. Drag the JSON into ComfyUI or use the load button.
3. Check every loader node — confirm the model paths match your folder structure. Filenames must be exact.
4. Load your reference images into the two VHS_LoadImagePath nodes. Use transparent background PNGs for character references for best results.
5. Load your reference video into VHS_LoadVideoPath.
6. Write your text prompt in the PrimitiveStringMultiline node. Use the subject_definitions format: Subject 1 (character from Picture 1), Subject 2 (environment from Picture 2), Video 1 (reference video motion).
7. Set the ResolutionSelector to your target aspect ratio. Default is 9:16 Portrait at 0.6 megapixels.
8. Click Queue. Watch Stage 1 output first before the full chain runs through SeedVR2, LTX, FILM, and RTX.
----------------------------------------
TIPS
----------------------------------------
1. SeedVR2 target resolution sits at 720 in this workflow. That is the sweet spot for edge sharpening without over-processing. Push it to 1080 if your GPU can handle it, but watch for over-sharpened artifacts on skin textures.
2. The reason SeedVR2 runs before LTX, not after: SeedVR2 sharpens edges but can introduce temporal flickering across frames. LTX 2.3 as a second pass smooths that flickering while keeping the edge detail. If you flip the order, LTX smooths first and SeedVR2 re-introduces flicker on top.
3. Turbo LoRA at 0.7 strength. The free workflow uses the same value. If motion looks under-driven, push to 0.85. Do not go to 1.0 — Turbo at full strength can produce artifacts on fast motion.
4. ResolutionSelector at 0.6 megapixels for Stage 1 is intentional. H3 generates at a lower resolution, then each subsequent stage scales up. Starting at a higher megapixel value does not help — it just uses more VRAM on Stage 1 without improving the final output.
5. Keep BasicScheduler at 4 steps with beta scheduler. Turbo LoRA is designed for 4-step sampling. Going above 8 steps wastes time with no visible improvement when Turbo is active.
6. The cache clear nodes (easy cleanGpuUsed, easy clearCacheAll) between stages are there for a reason. Each stage loads different models — H3 at 32B params, SeedVR2 at 7B, LTX at 22B. Without clearing, you will hit OOM on 24GB cards. Do not delete them.
7. RTX Video Super Resolution at ULTRA quality is the final pass. It requires an NVIDIA RTX GPU (20-series or newer). If you do not have an RTX card, bypass this node and use only the FILM interpolation output.
----------------------------------------
COMMON ISSUES
----------------------------------------
OOM during SeedVR2 stage — lower the SeedVR2 target resolution from 720 to 540. Also confirm the cache clear nodes between Stage 1 and Stage 2 are wired and active. SeedVR2 loads its own 7B DiT model on top of the H3 models still in VRAM.
Flickering in final output — check that LTX second pass (Stage 3) is not bypassed. LTX is what smooths the temporal flickering from SeedVR2. If you bypass LTX, the flicker passes through to the final output.
Faces still pixelated in mid or far shots — this is the core problem the multi-pass chain solves. Confirm SeedVR2 is running and its target resolution is at least 720. If faces are still soft after the full chain, raise SeedVR2 target to 1080 at the cost of longer processing time.
LTX second pass produces washed-out colors — this can happen if the color pipeline between SeedVR2 (lab color space) and LTX is mismatched. Check that the ImageScaleToMaxDimension node is set to lanczos at 1920 and is wired between SeedVR2 output and LTX input.
Frame rate mismatch in output — the workflow outputs at 24fps. FILM doubles the frame count, so if your Stage 1 output is 5 seconds at 24fps (120 frames), FILM produces 240 frames. The VHS_VideoCombine nodes are pre-configured to handle this — do not change the frame_rate values unless you understand the math.
SeedVR2 model not found — the SeedVR2 custom node pack is not in ComfyUI Manager by default. You need the ComfyUI-Seecoder or equivalent SeedVR2 node pack installed. Check that seedvr2_ema_7b_fp16.safetensors and ema_vae_fp16.safetensors are in the correct model folders.
----------------------------------------
Questions? Drop them in the comments and I will help you sort it out.
Thanks for supporting the channel. You guys make these deep dives possible.
— Benji