Create Video — MiniMax H3 (synchronized audio, ROCm)

Generate text-to-video and first-frame image-to-video with MiniMax H3 on the Radeon 8060S/gfx1151 through the persistent ComfyUI service at http://127.0.0.1:8188. MiniMax H3 is the only backend in this skill.

When to use

Load this skill unconditionally when the request contains 生成動畫, 生成影片, 製作動畫, 製作影片, 做動畫, 做影片, animate, animation, video, movie, film, short clip, or asks for MiniMax H3, Minimax H3, or Hailuo H3 moving-image generation.

Use it when:

• The user asks for a generated .mp4.
• The user provides a still and asks to animate it.
• The user wants synchronized ambience, SFX, dialogue, or music generated with the video.
• A longer story should be built from several polished 3–5 second H3 segments.
• A 2D anime request needs /create-image storyboard assistance before video generation.

Validated defaults

Setting	Default
Resolution	864×480
Frame rate	24 fps
Requested duration	5 seconds
Valid frames	124 (17k+5)
Actual duration	about 5.167 seconds
Sampler	Turbo v4 step600 EMA
Steps	6
Audio	native 32 kHz stereo, on
Typical wall time	about 5m25s

Use 8 Turbo steps when quality matters more than time. Use the 20-step reference sampler only for controlled comparisons or a shot that repeatedly fails to reach its endpoint.

Standard workflow

1. Convert the request into a shared cinematic Shot Map: story beat, shot size, viewpoint angle/height, simulated focal length, framing/depth of field, viewpoint movement, subject movement, continuity constraints, environmental progression, sound, and intended join.
2. For 2D anime, create horizontal storyboard/keyframe images with /create-image, passing --output-size equal to the planned H3 video resolution (normally 864x480).
3. Verify each PNG's pixel dimensions exactly match the video canvas, then show storyboard stills to the user and obtain visual approval before expensive generation.
4. Give each H3 segment one main character action and one environmental progression.
5. Generate one 3–5 second segment at a time.
6. Chain longer stories using motion context when available, otherwise the previous segment's actual final frame and optional FL2VA endpoint.
7. Assemble with duplicated-context trimming and a hard join or tiny audiovisual microfade.
8. Validate dimensions, fps, duration, video/audio streams, finite audio, visual continuity, and scene evolution before delivery.

CLI

$HOME/.pi/agent/skills/create-video/create_video.sh \
  -p "integrated_multimodal_description: [Shot 1] ... overall_soundscape: ... non_diegetic_music: ..." \
  -o output.mp4

Put an approved storyboard/reference image into MiniMax H3

The wrapper's --image option is the CLI input for an approved storyboard keyframe. It routes the image to H3 FL2VA as the exact first frame and geometry anchor; it is not a loose style reference. Use one clean horizontal frame generated at exactly the same resolution as the planned video—normally 864×480. A matching aspect ratio alone is not sufficient.

REF_IMAGE="/absolute/path/to/approved_storyboard_shot01_864x480.png"
OUT="$HOME/generated/minimax-h3/shot01.mp4"

$HOME/.pi/agent/skills/create-video/create_video.sh \
  --model minimax-h3 \
  --image "$REF_IMAGE" \
  --resolution 864x480 \
  --duration 5 \
  --steps 6 \
  --prompt "For the target video, at 0.00 seconds into the target video, Picture 1 is fully referenced. integrated_multimodal_description: [Shot 1] Preserve Picture 1's exact adult character identity, costume, horizontal composition, scene geography, landmarks, eye-level camera, and focal length. The character <one concrete action>. At the same time <one environmental change travels from foreground through mid-ground to background>. By the end <exact character and environment end state>. Traditional hand-drawn 2D cel animation; do not replace the background. overall_soundscape: <continuing ambience and action-linked sounds>. non_diegetic_music: <music development or none>." \
  --output "$OUT"

Paths containing spaces are safe when enclosed in double quotes. When creating the storyboard, set one shared resolution variable and pass the exact same value to both tools:

VIDEO_RESOLUTION="864x480"
STORYBOARD_DIR="$HOME/generated/minimax-h3/storyboards"
mkdir -p "$STORYBOARD_DIR"

cd "$HOME/models-work/flux2"
source .venv-rocm72/bin/activate
FLUX2_BIG_WMMA_LINEAR=1 python \
  "$HOME/.pi/agent/skills/create-image/scripts/create_image.py" \
  "Horizontal landscape 16:9 full-bleed anime image, not portrait or square. <exact shot design>" \
  --anime --aspect-ratio 16:9 --output-size "$VIDEO_RESOLUTION" \
  --out-dir "$STORYBOARD_DIR" --prefix shot01

# Confirm the returned final_path is exactly 864x480 before using it.
ffprobe -v error -select_streams v:0 -show_entries stream=width,height \
  -of csv=s=x:p=0 "/absolute/final_path/from/create-image.png"

$HOME/.pi/agent/skills/create-video/create_video.sh \
  --image "/absolute/final_path/from/create-image.png" \
  --resolution "$VIDEO_RESOLUTION" \
  --prompt "For the target video, at 0.00 seconds into the target video, Picture 1 is fully referenced. <motion, environment, and sound>" \
  --output "$HOME/generated/minimax-h3/shot01.mp4"

Do not generate a 1920×1080 storyboard for an 864×480 video and rely on H3 to resize it. Generate the storyboard at the final video canvas from the start. Use FFmpeg cropping only to repair legacy or externally supplied images that cannot be regenerated.

Direct ComfyUI generator equivalent, for debugging the wrapper:

cd "$HOME/src/ComfyUI"
./.venv/bin/python scripts/minimax_h3_generate.py \
  --image "/absolute/path/to/approved_storyboard_shot01_864x480.png" \
  --prompt "For the target video, at 0.00 seconds into the target video, Picture 1 is fully referenced. <observable motion, environment, and sound>" \
  --width 864 --height 480 --seconds 5 --steps 6 --seed 12345 --turbo \
  --prefix "video/storyboard_shot01" --server http://127.0.0.1:8188

Multiple storyboard rule: do not pass several unrelated storyboards into one wrapper call. Generate one H3 clip per approved storyboard. For a continuous segment 2, prefer clip 1's losslessly extracted actual final frame as segment 2's --image; use the next storyboard as a planned endpoint only in an explicitly endpoint-enabled FL2VA workflow. For a deliberate new viewpoint or focal length, use that storyboard as a new shot and join it with an editorial cut.

Options

Flag	Meaning	Default
--model	Compatibility selector: minimax-h3, minimax, or h3	minimax-h3
-p, --prompt	Observable visual and audio description	required
-i, --image	Exact first-frame image for FL2VA	none
-d, --duration	One-shot duration; 3–5 seconds recommended	5
--fps	Native fps; must remain 24	24
-a, --aspect	Aspect ratio when resolution is omitted	16:9
-r, --resolution	Explicit dimensions, snapped to multiples of 32	864x480
--hq	Use a 576-pixel short side	off
-o, --output	Output MP4	timestamped path
--seed	Random seed	random
--steps	Denoise steps	Turbo 6, reference 20
--turbo	Turbo v4 sampler	on
--no-turbo	Reference sampler	off
--audio	Keep synchronized native stereo audio	on
--no-audio	Strip audio from final MP4	off

Prompt structure

Use a structured multimodal prompt:

integrated_multimodal_description: [Shot 1] <subject, concrete action, environment,
lighting, viewpoint, focal length, framing, camera movement, temporal progression>.
overall_soundscape: <continuing ambience and synchronized visible SFX>.
non_diegetic_music: <genre, instruments, development, or "none">.

For image-to-video, begin with:

For the target video, at 0.00 seconds into the target video, Picture 1 is fully referenced.

Then specify what moves and changes. Do not redescribe a conflicting identity, costume, geometry, or viewpoint.

Cinematic Shot Map: angle, focal length, and movement

Before generating storyboards or clips, maintain one shared Shot Map as the source of truth. Do not invent independent image and video prompts: derive both from the same row.

Shot/beat | duration | shot size | viewpoint angle and height | simulated focal length
framing and depth of field | viewpoint movement | subject movement | start state | end state
environment and light progression | screen direction/continuity | audio | next-shot join

Shot size and viewpoint angle

Use shot size and viewpoint angle as separate controls:

• EWS/WS/full shot: geography, full-body action, choreography, and floor contact.
• MS/MCU: body language, dialogue, hand actions, and social tension.
• CU/ECU/insert: expression, eye line, impact reaction, or decisive prop detail.
• Eye-level: neutral human scale. Low angle: power or threat. High angle: vulnerability or overview.
• Overhead/bird's-eye: spatial geometry and choreography. Worm's-eye: extreme upward force, heroic jumps, falling debris, or vertigo.
• Dutch angle: instability or impact; state an observable horizon tilt, normally 10–25°, instead of merely saying “Dutch angle.”
• Profile/three-quarter/OTS/POV/reverse: specify which shoulder or character owns the foreground, where the view looks, and the subjects' spatial relationship.

Describe viewpoint geometry in natural language. Avoid the bare word camera in image prompts when a model may render equipment; use viewpoint, view, or framing. Video prompts may use standard cinematography terms when describing motion.

Simulated focal-length guide

Millimetres are visual intent for generative models, not guaranteed physical calibration. Always pair a focal length with visible perspective cues:

Simulated lens	Use	Observable cues to request
14–18 mm ultra-wide	worm's-eye, extreme speed, monumental space	strong near/far scale, enlarged foreground, converging lines, edge stretch
20–24 mm wide	close action, narrow interiors, movement toward view	pronounced depth, foreground energy, readable environment
28–35 mm moderate wide	establishing, full-body combat, tracking	natural wide perspective, character plus geography
40–50 mm normal	dialogue, medium shots, neutral observation	restrained distortion, natural proportions
65–85 mm short telephoto	portrait, emotion, close-up	shallow depth, clean face, separated background
100–135 mm telephoto	standoff, pursuit, surveillance	compressed distance, stacked background planes
200 mm+ long telephoto	remote observation, extreme compression	flattened depth, narrow field of view, strong isolation
Fisheye	comic aggression or surreal motion	curved edges and deliberate radial distortion

Do not choose lens by shot size alone: an 18 mm close-up exaggerates the face, while an 85 mm close-up flatters and isolates it. For character beauty shots, prefer 50–85 mm. For dynamic foreshortening, prefer 14–28 mm and explicitly identify the enlarged foreground limb or prop. Treat 28 mm establishing, 50 mm medium, and 85 mm close views as editorial cuts, not endpoints to be morphed together.

Viewpoint-movement grammar

For every moving shot specify movement type + direction/path + amplitude + speed + subject relationship + endpoint:

• slow push-in from a wide two-shot to a medium close-up, stopping before the face crops
• fast side-tracking left-to-right parallel to the runner, keeping the full body centered
• low 18 mm viewpoint retreats rapidly as the swordswoman lunges toward it
• crane rises three metres into an overhead view, revealing the circular spell pattern
• clockwise 120° orbit at constant radius while both fighters remain opposite each other
• whip-pan follows the strike and resolves on the opponent's impact pose

Available movement vocabulary includes static, pan, tilt, push/pull, tracking, side tracking, arc/orbit, crane/boom, pedestal, handheld, zoom, and dolly zoom. Distinguish viewpoint movement from subject movement. Avoid stacking more than one dominant viewpoint move and one subject action in a 3–5 second segment unless deliberately testing a complex shot.

Continuity checks

Track these across connected shots: identity, wardrobe, prop/weapon hand, eye line, screen direction, character facing, ground position, scene geography, landmark placement, light source direction, weather, environmental state, and action phase. Preserve the 180-degree rule unless a motivated crossing shot visibly re-establishes geography. Every shot must have a new story function; do not repeat the same pose, distance, and angle without narrative purpose.

For action, preserve readable causality:

anticipation/charge → initiation → contact or evasion → reaction → settle/new threat

A single storyboard frame should depict one readable instant, not several moments at once.

Segmentation and editorial assembly plan

Choose the join before generation; continuity does not mean forcing every clip through the previous final frame.

Relationship to next segment	First-frame source	Recommended join
Same continuous action and viewpoint	previous clip's actual final frame or motion context	remove duplicate context/frame; hard seamless join
Same scene, deliberate new angle or focal length	independently approved storyboard	cut-on-action, reaction cut, match cut, or audio J/L cut
New scene or time	new storyboard	motivated hard cut, ambience bridge, flash/occlusion if appropriate
Fast combat montage	independent 2–3 s shots	impact cut or whip-pan cut
Dialogue/emotional scene	independent 4–6 s shots	reaction cut with J/L-cut dialogue or ambience
Incompatible existing clips	dedicated bridge only if spatially plausible	otherwise keep an honest editorial cut

Use 3–5 second clips by default. Split whenever there is a major change in viewpoint, focal length, location, dominant action, or environmental state. For each segment write:

Segment N — duration/frame count; start state; one dominant action; one dominant viewpoint
move; environment/light progression; sound progression; exact end state; continuity source;
join type and trim/crossfade amount.

Assembly rules:

1. Prefer motion-context continuation for a truly continuous shot; otherwise extract the previous clip's lossless final frame for chained I2VA/FL2VA.
2. Use a fresh approved storyboard for a genuine cut. Do not let the previous frame lock a new angle into a weak morph.
3. Trim duplicated motion-context time or the repeated opening frame before concatenation.
4. Prefer a hard cut. Use only a 0.1–0.25 second audiovisual microfade when needed; never hide broken geometry with a long dissolve.
5. Preserve ambience across cuts with J/L cuts, but keep visible impact SFX synchronized.
6. Inspect several frames before and after every join for pose, velocity, screen direction, lighting, identity, audio phase, and background continuity.

MiniMax H3 operational rules

• Source images should match the target aspect ratio. For the tuned path, prepare the approved keyframe at exactly 864×480 because it is the first-frame geometry anchor.
• Width and height must be divisible by 32.
• Use native 24 fps.
• Valid frame counts follow 17k+5: 73 ≈ 3.04 s, 90 ≈ 3.75 s, 107 ≈ 4.46 s, and 124 ≈ 5.17 s.
• Keep one primary shot/action in each segment. Describe concrete movement rather than internal emotion.
• State dialogue, ambience, visible SFX, and music explicitly. Keep dialogue short and quote exact words.
• Do not enable global SageAttention; long-sequence backend corruption has been reported.
• Use the validated FP16 video VAE path. Do not force FP32 video VAE. Audio VAE stays FP32.
• This host is UMA. CPU offload does not release physical memory and only adds copies.
• Never stop or shrink qwen-mtp.service to make room. H3 is validated while it remains active.

Videos longer than five seconds

Never force a long request into one unreliable generation. Plan it as ordered 3–5 second H3 segments. Every segment needs:

Segment N — frame count; start state; character action; environmental progression;
camera/lens; sound development; exact end state; next context source.

Keep resolution, fps, identity, scene geography, screen direction, lens logic, and base art direction coherent across a continuous shot.

Preferred: latent and audio motion context

When a compatible H3 Motion Context workflow is installed, pass the previous segment's final video latent and audio context into the next generation. This preserves speed, direction, pose trajectory, color, ambience, and sound continuity better than decoding a single frame.

• Prefer 39 context frames (about 1.63 seconds) for routine continuation.
• Use 56 frames (about 2.33 seconds) for fast or complex movement when memory permits.
• Use 22 frames for calmer movement.
• Keep resolution unchanged throughout the chain.
• Continuation prompts must first preserve established movement, composition, environmental state, and sound bed, then introduce the next action.
• Trim duplicated context during assembly and inspect the exact join.
• The simple wrapper does not claim to perform latent/audio chaining automatically.

Community workflow: https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context

Portable fallback: chained I2VA/FL2VA

1. Generate segment A.
2. Extract A's actual final frame losslessly.
3. Use it as segment B's first_frame.
4. For free continuation, describe forward development. For a planned endpoint, use the next approved storyboard as last_frame and describe one continuous path toward it.
5. Remove the duplicated opening frame during assembly.
6. Prefer a hard join. If required, use only a 0.1–0.25 second audiovisual microfade.
7. Never use a long dissolve to hide incompatible poses; it produces ghosting.

If independent clips already exist, generate a separate 3–5 second FL2VA bridge from clip A's actual final frame to clip B's first frame.

Deliberate shot changes

Do not morph incompatible focal lengths or viewpoints. Use an editorial cut, cut-on-action, occlusion, petals, flash, or whip-pan. Preserve ambience with an audio J/L cut or tiny acrossfade. Frame interpolation can smooth cadence within a shot but cannot repair identity, pose, camera, or scene discontinuity.

2D anime storyboards with /create-image

For 二次元動畫, anime, cel-animation, and hand-drawn requests, use the create-image skill with Anima to design keyframes before H3 generation.

1. Define exact subject ground position and screen position, body orientation, foot placement, action, viewpoint side, viewpoint height, focal length, framing, fixed landmarks, and entry/exit movement for every shot.
2. Generate every storyboard horizontally and at the exact video resolution. Produce one clean, full-frame 16:9 landscape image per shot with /create-image --output-size 864x480 for the tuned default. If the video uses another canvas, pass that exact WxH to both create-image --output-size and create_video.sh --resolution. Never default to portrait, vertical, square, panels, or a contact sheet.
3. Begin image prompts with: Horizontal landscape 16:9 full-bleed image, not portrait or square.
4. Do not put panel numbers, captions, lens labels, camera diagrams, borders, or fake UI into image prompts.
5. Use non-native draft generation first. Spend native 1080p time only after art-direction approval or when the user explicitly requests a final high-quality still.
6. For hand-drawn style, specify varied ink lines, flat opaque cel colors, hard two-tone shadows, and painted gouache/paper backgrounds. Exclude 3D, CGI, glossy rendering, plastic skin, bloom, bokeh, volumetric effects, visible camera equipment, text, and watermarks.
7. Establish a character master and scene map. Reuse appearance, palette, screen direction, ground markers, and landmarks. Use conservative img2img for adjacent poses and inspect identity, hands, feet, clothing, background, and perspective after every generation.
8. Show stills to the user and obtain visual approval before generating video.
9. Use same-composition adjacent keyframes as FL2VA endpoints. Treat 28 mm establishing, 50 mm medium, and 85 mm close views as deliberate cuts rather than forced morphs.
10. Confirm the create-image JSON final_path has the exact target dimensions before video generation. Do not silently resize a newly generated storyboard afterward; matching pixel geometry must be established during storyboard generation.

Image prompt skeleton:

Horizontal landscape 16:9 full-bleed anime image, not portrait or square.
<adult character design> stands <exact world position> and appears <screen position>.
Body faces <direction>; feet and weight <placement>. Viewpoint is <side and height>,
<focal length>, <shot size>. <Fixed landmarks and screen direction>.
Traditional hand-drawn 2D animation keyframe, varied ink lines, flat cel colors,
hard two-tone shadows, painted paper background. No panels, captions, text, visible
camera, 3D, CGI, glossy rendering, bloom, bokeh, or watermark.

Animation philosophy: animate the world, not only the character

A successful animation is not a moving character pasted over a frozen illustration. Treat the character, environment, light, atmosphere, sound, and camera as one causal system. A static camera means stable framing, not a static world.

1. Protect invariants; animate variables. Keep identity, costume, geography, landmarks, screen direction, lens, and art direction stable. Deliberately evolve wind, grass, branches, clouds, water, crowds, practical lights, shadows, weather, particles, or distant activity.
2. Give the environment an arc. Every segment needs an environmental beginning, progression, and readable ending. The ending becomes the next segment's starting state.
3. Use visible cause and effect. A spell sends a light wave across flowers; footsteps disturb dust and grass; a door changes interior light. Avoid unrelated decorative motion.
4. Stage change through depth and time. Describe when motion reaches foreground, mid-ground, background, and landmarks. Use begins, travels, one after another, gradually, and settles rather than listing simultaneous effects.
5. Keep change legible. Prefer one dominant environmental transformation plus two or three supporting motions. Too many unrelated changes cause geometry replacement and flicker.
6. Preserve geography during transformation. Progressively transform the established scene; never suddenly replace the background.
7. Let lighting tell time and emotion. Cloud shadows, dusk shifts, lanterns, reflections, and magic illumination must affect both character and environment consistently.
8. Make sound undergo the same event. Continue ambience and add spatial cues in causal order. Music may develop with the transformation but must not restart as an unrelated take.
9. Use stillness intentionally. Held poses may contrast action, while breathing, cloth, light, and atmosphere retain life.
10. Judge continuity as motion, not matching pixels. Inspect velocity, direction, illumination, sound, and environmental state around every join.

Temporal prompt pattern:

At first <stable character and environmental state>.
Then <character action> causes <dominant environmental event>.
The effect travels from <foreground> through <mid-ground> to <background landmark>.
Meanwhile <two supporting motions> evolve consistently.
By the end <character pose, lighting, environment, and sound settle into next state>.
Preserve <identity, geography, camera, lens, and art style>; do not replace the background.

Deployment and performance

• ComfyUI: ~/src/ComfyUI
• Service: ~/.config/systemd/user/comfyui.service
• API: http://127.0.0.1:8188
• Generator: ~/src/ComfyUI/scripts/minimax_h3_generate.py
• Diffusion model: minimax_h3_fl2va_pruned_int8_convrot.safetensors
• Text encoder: qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
• Video VAE: minimax_h3_video_vae_fp16.safetensors
• Audio VAE: minimax_h3_audio_vae_fp32.safetensors
• Turbo LoRA: minimax_h3_turbo_v4_step600_ema.safetensors

The gfx1151 optimization materializes contiguous Q/K/V before SDPA. On the tuned workload it improved sampling from about 461.6 seconds to 222 seconds and end-to-end wall time from about 565.4 seconds to 325.2 seconds. Do not remove this optimization without an exact-shape A/B.

Validation before delivery

ffprobe -v error \
  -show_entries format=duration,size \
  -show_entries stream=index,codec_name,codec_type,width,height,r_frame_rate,sample_rate,channels \
  -of json output.mp4

Confirm:

• MP4 exists and is non-empty.
• Video is H.264, requested dimensions, and 24 fps.
• Native-audio output contains AAC 32 kHz stereo.
• Decoded audio contains no NaN or Inf.
• Beginning, middle, end, and every segment join are visually inspected.
• Character identity, hands, feet, costume, geography, and screen direction remain coherent.
• The environment visibly evolves rather than behaving like a frozen illustration.
• Sound development matches visible events.
• No long ghosting dissolve hides a broken join.
• comfyui.service and protected qwen-mtp.service remain healthy.

Sources

• Official ComfyUI workflow: https://docs.comfy.org/tutorials/video/minimax/minimax-h3
• Official model repository: https://huggingface.co/Comfy-Org/MiniMax-H3
• Official prompt guide: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md