Pika H3 Flex 3-Sampling Workflows: NVRegen 2K and RTX Refine - For Patreon Supporters (Patreon)
Videos
-
demo shorts PikaH3.mp4 -
776441864.mp4
Downloads
Content
Hey everyone. If you watched the video you saw the A/B tests and the upscale results.
And The Youtube video : https://youtu.be/5sC-FjwGqmQ
This post has the two production workflows I kept running after the tests: the OpenH3 Regenerate 2K variant and the RTX refine variant, with the exact strengths, step counts and schedulers from my runs. The free A/B workflow stays in the public post; these two are the ones with the extra sampling stages.
Demo:
----------------------------------------
WHAT TO DOWNLOAD (MODEL FILES)
----------------------------------------
Base and encoders:
- minimax-h3/Minimax-h3_Singularity_ref2va_v2_int8.safetensors (huggingface.co/WarmBloodAban/Minimax-h3_Singularity)
- minimax-h3/qwen3vl_32b_minimax_h3_int8_convrot.comfy.safetensors (text encoder, CLIPLoader type minimax)
- minimax-h3/minimax_h3_video_vae_fp16.safetensors (frames)
- minimax-h3/minimax_h3_audio_vae_fp32.safetensors (soundtrack)
LoRAs:
- pika_h3_flex/pika_h3_flex_turbo_8step_hd_ref2va.safetensors (huggingface.co/akhaliq/Pika-H3-Flex-Turbo-LoRA)
- minimax-h3/h3-realism-people-t2v-i2v-r2v(r34l1sm).safetensors
- minimax-h3/minimax_h3_lms_v1.0_r64.safetensors (refine and upscale stages)
- Open-H3-Regenerate-2K/sharp_curve_branches.safetensors (variant A upscale stage only)
- Open-H3-Regenerate-2K/open_h3_regenerate_2k_lora.safetensors (variant A only)
- minimax-h3/minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors (only if you also run the free A/B workflow; the two supporter variants no longer load it)
Custom nodes:
- ComfyUI 0.30 or newer (native MiniMax H3 reference nodes)
- ComfyUI-VideoHelperSuite (video combine and image path loaders)
- ComfyUI-KJNodes (ImageResizeKJv2 for reference normalization)
- ComfyUI-Easy-Use (the cleanGpuUsed and clearCacheAll maintenance nodes)
- OpenH3 Regenerate node pack (variant A upscale stage)
- ComfyUI-NVIDIA-RTX-VFX for frame generation and super resolution (variant B)
- DLSS 5 enhance nodes (both variants)
- An RMBG background-removal node for the reference cutouts
- A sparse attention pack (BlockSparseAttention in both variants)
An NVIDIA RTX GPU only matters for variant B. Variant A runs fine on any card that handles the int8 base.
----------------------------------------
QUICK START
----------------------------------------
1. Load the variant you want and let ComfyUI resolve the nodes.
2. Check the four loader paths: UNET, text encoder, video VAE, audio VAE. Then check the loras subfolder names.
3. Drop your reference images into the LoadImage nodes. The RMBG node cuts the background and the resize node normalizes them to 1230x1024 before conditioning.
4. Paste your prompt into the text node (format below).
5. Set your duration and resolution on the MiniMaxH3ReferenceToVideo node. Generation runs at 1344x768.
6. Fix a seed while iterating, randomize when you like the prompt. Queue it.
That's the whole loop. The refine and upscale stages are wired after stage 1 and run on their own.
----------------------------------------
WHAT TO INPUT (PROMPT FORMAT)
----------------------------------------
These are reference-to-video runs, so use the six-part Ref2VA form. Label every reference, keep one dominant action per shot, give every shot a camera move, and always write a sound line. Short free-form prompts render noticeably worse on this model.
subject_definitions:
<Subject 1> is the woman in <Picture 1>, silver-lavender undercut hair, cropped orange bomber jacket, red and gold armored limbs.
<Subject 2> is the woman in <Picture 2>, hime-cut black hair, round glasses, black turtleneck, pale-blue cybernetic legs.
summary:
[reference generation] One paragraph: who does what, where, in how many shots.
detailed_description:
One or two style sentences first, then [Shot 1] with no timestamp, then [Shot 2] At 00:04.000 and so on. Each shot: framing, one action, camera move with amplitude and speed, what the light does.
overall_soundscape:
Room tone plus material sounds for the whole clip.
non_diegetic_music:
N/A, or one sentence of score the characters cannot hear.
Dialogue goes inside <d> tags with a language label: <Subject 2> (S1) says in a quiet voice, <d>[English] The city's quieter from up here.</d>
----------------------------------------
KEY SETTINGS (AND WHY)
----------------------------------------
Both variants, stage 1:
- Plain base model with the realism LoRA at 0.6. Euler with beta scheduler, 12 steps, denoise 1. No turbo LoRA on this stage.
- Why: the high-sigma steps lay down structure and noise. A distilled LoRA on this stage fights the noise ladder, so the Pika file waits for the refine stage.
Variant A, Minimax-Pika-H3_3sampleing-NVRegen2k-20261102.json (109 nodes):
- Stage 2, split-sigma refine: Pika 8-step HD LoRA at 1.0 on the same realism 0.6 chain, euler.
- Stage 3, OpenH3 Regenerate 2K upscale: Regenerate 2K LoRA at 1.0, sharp curve branches at 1.0 and LMS at 0.8 stacked on the same chain, euler.
Variant B, Minimax-Pika-H3_3sampleing-RTXRefine-20261102.json (109 nodes):
- Refine stage: Pika 8-step HD at 1.0 and LMS at 0.8 on the same chain, euler with simple scheduler, 4 steps, denoise 0.3.
- Upscale: MMH3 latent upscale to 1280x704, then RTX frame generation, RTX super resolution and the DLSS 5 enhance finish.
- Why denoise 0.3 on the refine pass: 0.5 re-melts faces, 0.3 sharpens without rewriting them.
Both variants: attention backend, sparse attention and the cache nodes ship correct. Leave them as shipped.
----------------------------------------
TIPS
----------------------------------------
1. Never stack a second turbo LoRA on top of the Pika file. It is already distilled; a second acceleration LoRA breaks the output. I show the broken render in the video.
2. Feed clean cutouts. The RMBG node handles busy backgrounds, but a reference with a flat backdrop gives the conditioning more to work with.
3. The 6+6 split sigma run in the video was the test bench. The shipped workflows settle on 12 plain steps in stage 1 plus a short refine stage, so start there before rebuilding the split by hand.
4. Keep at least 4 steps on any turbo route if audio matters. Below about 6 steps the soundtrack degrades, and 2-step files are draft-only.
5. For high action, use variant A or B rather than a single sampler. One sampler on a fight scene gives you mush; the extra stages keep the contacts readable.
6. Pick the LoRA per scene, not per project: Pika for far and mid shots with faces, LightX2V for talking scenes where hand motion carries the performance. The free A/B workflow is the one that still loads LightX2V.
7. If you want the Regenerate 2K look but the faces go soft, variant B's MMH3 latent upscale plus DLSS 5 is the cleaner path. My retake comparison has the plain latent upscale winning on color and hair.
----------------------------------------
COMMON ISSUES
----------------------------------------
Pixelated faces in far shots on the LightX2V route: that is the known first-sampler limit. Add the second high-res sampling pass, which is exactly what variant B's refine stage does.
Motion looks overcooked or waxy: check that the Pika LoRA sits on the refine stage, not stage 1. A distilled LoRA on the high-sigma stage fights the noise ladder; stage 1 wants the plain base with realism at 0.6.
Upscale stage fights the Pika turbo, smeared faces or broken armor plating: bypass open_h3_regenerate_2k_lora in variant A and keep sharp curve branches plus LMS, or switch to variant B. That matches the retake: the plain latent upscale held color and hair better than the 2K pass.
VRAM crashes at 1344x768: the cache-clearing nodes are already wired. Drop the RTX and DLSS stages first since they hold the upscaled frames, then shorten the clip and render in passes.
Clip length comes out slightly longer than requested: durations round up to the model's frame grid. Fifteen seconds is 362 frames at 24 fps. That is normal.
----------------------------------------
Both workflow files are attached. Start with variant A if you want the Regenerate 2K look, or variant B if you have an RTX card and want the DLSS finish.
Questions? Drop a comment with your GPU, VRAM, and the exact error and I'll help you sort it out.
Thanks for supporting the channel. You guys make these deep dives possible.
- Benji