MiniMax H3 Ref2VA + Upres to 4K — Patreon Exclusive (Patreon)
Videos
-
2026-08-06 02-38-41.mp4 -
480 1080 after upsampling.mp4
Downloads
Content
Hey everyone,
The open-source MiniMax H3 Base only ships the low-res pass. Official 2K sits in H3-Regenerate-2K, and MiniMax still has that module closed. They say they will release it when ready. Until then the full 2K path runs through their API.
The last post about this AI video model : https://www.patreon.com/aifuturetech/posts/minimax-h3-local-165674768
Here is the problem.
The company said "Due to the complexity of the system, this module is not yet open-sourced. We will release it once it is ready. We provide an API for validating the official results; see "Full 2K Workflow" below."Well, those "complexity" are beautiful excuse. Basically, they just don't want freebies to get good resolution output from the generation.
I built this local workaround: generate with H3 Base Ref2VA at 480p or 720p, then run the clip through an LTX 2.3 second pass, FILM frame interpolation, and RTX Video Super Resolution up to 4K. Same character references. Same audio path. You stay local while H3-Regenerate-2K stays locked..
One file attached:
1. minimax_h3_r2v-ltxupres.json — full H3 Ref2VA + LTX upres + VFI + RTX 4K pipeline
H3-Base tops out around 768p-class (shorter side ~768 in the official docs). H3-Regenerate-2K is MiniMax's in-context regen to 2K, and that module is closed. This stack takes the open Base Ref2VA clip and rebuilds sharpness with LTX + RTX. It is not a copy of MiniMax's regenerate path. It is the local route that keeps you producing at high res now.
----------------------------------------
WHAT YOU NEED
----------------------------------------
MiniMax H3 (Stage 1):
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
Download H3 Comfy weights from:
https://huggingface.co/Comfy-Org/MiniMax-H3
Official model card / 2K status note:
https://huggingface.co/MiniMaxAI/MiniMax-H3
LTX 2.3 upres (Stage 2):
models/checkpoints/ltx2/ltx-2.3-22b-dev.safetensors
models/text_encoders/gemma_3_12B_it_fp8_scaled.safetensors
models/vae/LTX23_video_vae_bf16.safetensors
models/loras/ltx2/ltx-2.3-22b-distilled-lora-384.safetensors
models/loras/ltx2/LTX2.3_Crisp_Enhance.safetensors
models/loras/ltx2/ltx-2-19b-ic-lora-detailer.safetensors
LTX 2.3 checkpoint:
https://huggingface.co/Lightricks/LTX-2.3
----------------------------------------
TIPS
----------------------------------------
Start at 0.4 megapixels (≈864×480) for fast Ref2VA tests. Once the prompt and character lock look right, move to 0.7–1.0 MP for a cleaner 720p-class base before the LTX pass. Skip max size on H3 while you still iterate prompts. H3 Ref2VA burns VRAM and punishes vague wording.
Tag references the official way: <Picture 1>, <Picture 2>, matching the order of your Load Image nodes. Name which reference drives face, outfit, or motion. Vague prompts make Ref2VA drift.
ImageScaleToMaxDimension at 1920 is the handoff into LTX. The graph note: 1920 for roughly 1920×1080, 1280 for 1280×704. If the LTX pass OOMs, drop that max dim to 1280 before you touch LoRA strengths.
Keep Crisp Enhance and the Detailer IC-LoRA around 0.6. Higher values crunch skin and fabric after RTX. Distilled LoRA at 1.0 matches the short 4-step second pass.
Guiding strength 0.5 on LTXVLoopingSampler follows H3 motion and still leaves room for LTX detail. Face warp on the second pass? Raise guiding. Soft copy of the low-res frame? Lower guiding or push the Detailer LoRA.
RTX ULTRA ×2 after FILM ×2 gets you near 4K from a 1080-class LTX output. Save the H3 base and LTX mid outputs so you can A/B without re-running the full chain.
Use the clear GPU cache nodes between stages. Run Stage 1, check the preview, then run Stage 2/3. On tighter cards, generate H3 first, unload, then run the LTX+RTX half.
----------------------------------------
COMMON ISSUES
----------------------------------------
Soft ~480p/720p with no 2K look — H3 Base alone stops there. H3-Regenerate-2K is not in the open weights. Run this LTX+RTX chain, or use MiniMax's API full 2K workflow.
Character identity drifts after LTX — guiding strength too low, or ImageScaleToMaxDimension stretching a soft H3 frame. Raise guiding toward 0.6–0.7, or regenerate Stage 1 at a higher megapixel so LTX has a cleaner face to follow.
Face / fabric over-sharpened at 4K — drop Crisp Enhance or Detailer from 0.6 toward 0.4, or set RTX quality below ULTRA.
Out of VRAM on LTX second pass — lower ImageScaleToMaxDimension from 1920 to 1280, shorten duration, or disable EasyCache / Sage attention one at a time. The 22B LTX pass is the heavy step.
H3 length / frame errors — use the Duration primitive and the Math Expression node in the graph. Skip hand-typed frame counts. The expression rounds to a valid 24fps length for H3.
Audio out of sync after VFI — leave trim_to_audio off unless you want picture cut to sound. Keep src fps at 24 into the LTX condition and combines. FILM doubles frames; the combine nodes need that matching rate path.
Missing MiniMaxH3ReferenceToVideo node — update ComfyUI. H3 support landed via the Comfy-Org PR. Desktop/Cloud stable builds lag nightly.
Model filename mismatch — this graph loads minimax_h3_ref2va_int8_convrot.safetensors and the int8_convrot Qwen3-VL CLIP. If you grabbed pruned or nvfp4 variants from older notes, point the loaders at the files on disk.
----------------------------------------
Download the H3 and LTX models from the links above before you run it. This local bridge holds until MiniMax open-sources H3-Regenerate-2K. We can compare then.
Drop questions in the comments.
Thanks for supporting the channel.
— Benji
Have fun :)