Innovate Futures @ Benji

H3 Action Scenes Update: Singularity 3-Sampling Workflow + Character Sheets - For Patreon Supporters (Patreon)

Published:

2026-09-28 13:15:00

Edited:

2026-09-28 13:43:32

Imported:

2026-09

Videos

753720223.mp4

Downloads

Content

Hey everyone!

If you watched the H3 Action Scenes Update video (https://www.patreon.com/aifuturetech/posts/ai-video-minimax-170794297) , you saw the split — the old two-pass sampling with latent upscale, and the new multi-sampling setup on the same Singularity model. The free post has the beginner workflow. This one has the full 3-sampling version I tested through the whole video, plus the Qwen Image 2.1 character sheet workflow the fighters came from. Everything below is how to actually run it.

----------------------------------------

WHAT TO DOWNLOAD (MODEL FILES)

----------------------------------------

Base model (loaders point at a minimax-h3 subfolder, adjust to your paths):

- Minimax-h3_Singularity_ref2va_v1.3_int8.safetensors — the fine-tune, everything runs on this

https://huggingface.co/WarmBloodAban/Minimax-h3_Singularity

Text encoder and VAEs (Comfy-Org pack):

- qwen3vl_32b_h3_ultra_uncensored_heretic_int8_convrot.safetensors (or the standard qwen3vl_32b_minimax_h3 from the Comfy-Org text_encoders folder)

- minimax_h3_video_vae_fp16.safetensors

- minimax_h3_audio_vae_fp32.safetensors

https://huggingface.co/Comfy-Org/MiniMax-H3/

LoRAs (all three are wired into the workflow):

- minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors — strength 1.0, runs on the base model

https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/loras/minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors

- h3-realism-people-t2v-i2v-r2v(r34l1sm).safetensors — strength 0.65, refiner pass only

- https://huggingface.co/JOKER141/MiniMax-H3-Combat-Base-V2

- https://huggingface.co/JOKER141/MiniMax-H3-General-Motion-Continuity-Repair

- minimax_h3_lms_v1.0_r64.safetensors — strength 0.8, refiner pass only, this is the sharpening LoRA

https://huggingface.co/RunningHubAI/rh-minimax-h3-lms-v1.0-r64-lora

Latent upscaler:

- h3_upscaler_lms_v0.1_fp32.safetensors

https://huggingface.co/Alissonerdx/Minimax-H3-ComfyUI/blob/main/latent_upscaler/h3_upscaler_lms_v0.1.safetensors

Optional (end of the graph):

- film_net_fp16.safetensors for frame interpolation

- DLSS 5 enhance nodes if you have an NVIDIA card

For the character sheet workflow (QwenImg_2-1_t2i-charactersheet.json):

- qwen-image-2_1_bf16.safetensors

- qwen3.5_9b_qwen_image_2.1_pe_t2i.int8_convrot.safetensors (or qwen3vl_8b_bf16.safetensors)

- qwen_image_2.1_vae_bf16.safetensors

----------------------------------------

QUICK START

----------------------------------------

1. Load MinimaxH3_Singularity-3Sampling.json.

2. Check the loader paths — UNETLoader, CLIPLoader, both VAELoaders, and the three LoRA loaders. All of mine point at a minimax-h3 folder, so fix those first or nothing runs.

3. Drop your reference images into the VHS_LoadImagePath nodes in the References group. One image for a single-character shot, up to four for a multi-character fight.

4. Paste your prompt into the multiline string node in the Setting group.

5. Set the Duration float (seconds) and the ResolutionSelector megapixels. Seed is on the RandomNoise node.

6. Queue it. That's the whole loop.

The free beginner workflow in the public post folder is the same idea with the extra stages stripped out — start there if you just want to see it move.

----------------------------------------

WHAT TO INPUT (PROMPT FORMAT)

----------------------------------------

This is a reference-to-video workflow, so the prompt has two jobs: define who is on screen, then describe the shot. The structure the model likes:

subject_definitions:

<Subject 1> is the crimson ninja from <Picture 1>: an adult woman in a dark crimson bodysuit with twin sai in reverse grip, low predatory crouch, explosive and elastic fighting style.

<Subject 2> is the ice empress from <Picture 2>: an adult woman with a frost war-fan and ice magic, regal stance, sharp deliberate counters.

<Subject 3> is the location: the storm temple rooftop, wet dark stone tiles, lightning-lit clouds behind the skyline.

<Picture 1> is the identity and wardrobe reference for <Subject 1>; do not reproduce its sheet layout or background.

<Picture 2> is the identity and wardrobe reference for <Subject 2>; do not reproduce its sheet layout or background.

<Picture 3> is the location reference for <Subject 3>.

scene:

[12s] Wide establishing angle. The ninja sprints in low from frame left while the empress raises her fan and the rooftop frosting spreads. [2s] The ninja closes distance with a springing leap. [3s] Double-blade cross-cut meets a wall of ice, sparks and shards burst outward. [2s] The camera whips around them as the empress counters.

Rules that matter:

- Number your subjects in order and match the Picture labels to the exact image slots you plugged in (ref_image_1 is <Picture 1>, and so on).

- Describe the location as its own subject when you have a location image. The model holds the arena much better that way.

- Keep the shot block in seconds. Fast action needs short, specific beats — one action per sentence, named camera move when you want one.

- You do not need any trigger words. The Singularity model reads plain descriptive prompts.

If you feed the same idea to an AI agent with a prompt-writing skill built from the Singularity prompt spec, it formats all of this for you. That is how the fight prompts in the video were written.

----------------------------------------

KEY SETTINGS (AND WHY)

----------------------------------------

Sampling (this is the part that changed):

- BasicScheduler: beta, 6 steps, denoise 1.0. Beta is what the Singularity setup wants. Leave denoise at 1.0.

- ExtendIntermediateSigmas: steps 2, start_at_sigma 1.0, end_at_sigma 0, linear. This inserts extra sigma values between the scheduler steps, which is where the extra detail comes from. More extend steps means finer detail and a much longer render. This value is what the Singularity author recommends.

- SplitSigmas (first one): step 2. This gives the first sampler just 2 steps — it only lays down the latent noise foundation at low resolution. Everything from that split onwards goes to the later passes. If the first pass output looks like pure noise, that is correct.

Resolution and duration:

- ResolutionSelector: 0.6 megapixels 16:9 as shipped (multiple 32). The first pass resolution barely matters — I ran 0.5 in the video and the second pass still gave me the big clean output. The note in the graph has the full megapixel-to-size table.

- MinimaxH3LatentUpscaler3D: h3_upscaler_lms_v0.1_fp32, scale by multiplier, 1.2 as shipped. This is the scale that feeds the node from the Float node, so change that value, not the field on the node. Push it toward 1.5 when you want the biggest output like the video demos.

- Duration float is in seconds. The math node converts it to frames at 24 fps and rounds to the length the model likes (10 seconds gives 243 frames, 15 seconds gives 362). Just set seconds.

LoRA strengths:

- Turbo 4-step LoRA: 1.0 on the base model. This is the low-step setup the Singularity author pairs with the fine-tune.

- Realism LoRA: 0.65, refiner pass only. This is the anti-plastic one.

- LMS LoRA: 0.8, refiner pass only. This is the sharpening pass. If skin starts looking plasticky, drop this one to 0.4-0.7 before touching anything else.

Patches:

- Sage Attention patch on auto, memory-efficient patch on, chunked feed forward at 4096, low-VRAM attention on. That is the whole stack and it ships working. The low-VRAM helpers are optional on bigger cards.

Output:

- Frame interpolation x2 (film_net) then DLSS 5 enhance, final combine at 48 fps. Bypass the DLSS 5 stage when you want to see what the model alone is doing.

----------------------------------------

TIPS

----------------------------------------

1. Do not chase the first-pass resolution. The second and third samplers decide what you get. Low first pass, big second pass is the whole trick here.

2. The split sigma setup is slow on purpose. The normal two-pass method finishes a 15-second clip in about 2-3 minutes; this one takes a lot longer because the later passes are running at higher resolution with more steps.

3. First generation after loading always feels slow because everything gets loaded and patched. The next ones are faster.

4. Plasticky skin or waxy faces: the Realism LoRA is already in at 0.65. Raise it a little, or lower the LMS LoRA. Do not raise both at once or you lose the sharpness you added it for.

5. Water, dust and hair are the easiest way to see the difference between the two methods. Prompt some in on purpose when you are testing settings.

6. Keep the graph minimal. You do not need a pile of low-VRAM nodes to run this — only add the chunk helpers if you actually hit memory limits.

7. If you need production-grade sharpness, upscale again after the fact. I use Topaz for that, or the NVIDIA RTX upscaler for quick jobs.

8. Generate the character sheets first, then fight with them. Two or three reference images plus a location image gives the most stable multi-character action.

----------------------------------------

COMMON ISSUES

----------------------------------------

Fast motion comes out blurred or pixelated on the two-pass method -> that is the limitation this workflow fixes. Use the split sigma path and let the second pass carry the remaining steps.

Hair, water bubbles or rocks look mushy in the final output -> the refiner passes are not doing their job. Check that the LMS LoRA and the realism LoRA are actually connected to the second sampler model, not the base.

Skin looks plasticky -> lower the LMS LoRA to 0.4-0.7, or push the realism LoRA up slightly.

First sampler output is just noise and nothing else -> normal. That latent noise is the foundation for the later passes. Do not judge the result until the last sampler finishes.

Generation takes forever -> expected when the upscale multiplier and extend steps are both high. Drop the multiplier toward 1.0 first, then the extend steps.

Movement cuts off before the action finishes -> your Duration is too short for the prompt. The shot block is in seconds and the frame math follows it — if the prompt describes 15 seconds of action, set 15 seconds.

Sound comes out cleaner than expected compared to the old method -> that is real, the multi-sampler output carried crisper audio in every test I ran.

Questions or something weird in your output? Drop a comment with your GPU, VRAM, and the exact error and I will help you sort it.

Thanks for supporting the channel. You guys make these deep dives possible.

— Benji

----------------------------------------

REFERENCE LINKS

----------------------------------------

Diffusion, Text Encoder, VAE

https://huggingface.co/Comfy-Org/MiniMax-H3/

minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors

https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/loras/minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors

minimax-h3-lms-v1.0-r64-lora