H3 · self-hosted

H3 video model, self-hosted on Wavemaker's GPU fleet. Pick it in the model picker or name it in your brief. This page gathers its model configurations and example graphs.

Discover · Wavemaker catalog

When to use it

  • H3 video model, self-hosted on Wavemaker's GPU fleet. Pick it in the model picker or name it in your brief. This page gathers its model configurations and example graphs.

What you provide

  • Owned first and last frames, plus a brief for the transition.
  • Owned identity/reference images and a scene/motion brief.
  • A brief and the image, video or audio references needed for the selected task.
  • An owned starting image and a scene brief for both passes.

What you get

  • An editable configuration and bound Comfy graph; rendered media is produced only after a separate run.
  • An explicit h3 model configuration and example graph.
  • An editable Comfy graph with the stages listed in Technical details. Generated outputs depend on the selected runtime and bound inputs.
  • Output stages: Video.

Creative guidance

  • H3 first/last-frame video · Acc-8 study: Evaluate the Acc-8 adapter on a start/end-frame transition; its recorded trial executed but failed sampled-frame creative acceptance.
  • H3 reference video · Acc-8 study: Evaluate the Acc-8 adapter on identity-reference video; its recorded trial executed but failed sampled-frame creative acceptance.
  • h3 audio transplant: h3 model configuration for text-to-video.
  • h3 audio transplant i2v: h3 model configuration for image-to-video.
  • h3 chain window: h3 model configuration for first/last-frame video.
  • h3 dense native: h3 model configuration for first/last-frame video.
  • H3 · draft and refinement comparison: Retain separate audiovisual draft and refined outputs from the two-pass graph; the recorded trial did not demonstrate improved product fidelity.
  • H3 video · dense B200 service: Select the dedicated B200 service with explicit dense 20-step inputs; this preserves dense execution despite Turbo defaults in the retained example.
  • H3 video · dense model service: Use the explicitly configured dense H3 model-server route instead of loading the full model inside this graph.
  • H3 reference video · model service: Send identity references to the separately configured H3 reference-model service.
  • H3 video · Turbo model service: Select the dedicated Turbo model-server route; the retained two-shot text-to-video trial does not qualify every listed task.
  • h3 extend: h3 model configuration for video-to-video.
  • h3 face refine: h3 model configuration for video-to-video.
  • H3 text video · FastH3 preview: Explore the four-step FastH3 text-video preview route on its dedicated H100 service; it is a separate model choice; no execution or output qualification is recorded.
  • h3 fl bsa: h3 model configuration for first/last-frame video.
  • H3 first/last-frame video · native: Connect start and end frames with the 20-step FL2VA configuration; select the separate task-specific example when using text or one image.
  • h3 fl turbo4: h3 model configuration for text-to-video, image-to-video, first/last-frame video.
  • H3 first/last-frame video · Lightning 8-step: Use the eight-step Lightning adapter for first/last-frame video; this is a different adapter and conditioning path from reference video.
  • h3 long take: h3 model configuration for image-to-video.
  • h3 long take speech: h3 model configuration for image-to-video.
  • h3 motion pose: h3 model configuration for v2v recast.
  • h3 motion prompt: h3 model configuration for v2v recast.
  • h3 motion viggle: h3 model configuration for v2v recast.
  • H3 image animation · native Turbo: Animate one source image with the exact 768p Turbo LoRA and eight-step native graph.
  • h3 person swap: h3 model configuration for v2v recast.
  • h3 recamera crossview: h3 model configuration for v2v recamera.
  • h3 ref bsa: h3 model configuration for reference-to-video.
  • h3 ref fast: h3 model configuration for reference-to-video.
  • H3 reference video · native: Use several identity reference images with the 20-step reference-to-video configuration.
  • H3 reference video · SageAttention: Use the reference-video graph with the guider-only SageAttention FP16 patch; the scheduler keeps the original model and a compatible CUDA wheel is required.
  • h3 ref turbo4: h3 model configuration for reference-to-video.
  • h3 turbo i2v 1088: h3 model configuration for image-to-video.
  • h3 turbo i2v 1088 speech: h3 model configuration for image-to-video.
  • h3 turbo t2v 1088: h3 model configuration for text-to-video.
  • h3 window latent: h3 model configuration for video-to-video.
  • h3 window latent first: h3 model configuration for image-to-video.
  • Cartoon clip to live action: Attach a short animated clip and restyle it as live action. Up to about eight seconds — not a full episode or film. Gemini restage of first and last frames is billed.
  • Cartoon film window: One Gemini first-and-last restage plus an H3 video hop for a packed cartoon-film window. Gemini is billed. Member recipe for cartoon film to live action — not a Discover starter.
  • Model · MiniMax: MiniMax H3 self-host stereo 768p first, Draft 480p, then Hailuo hosted 2K (moderated), plus Speech.
  • Model · Self-hosted: WAN, Hunyuan, Foley, LTX, and MiniMax H3 on our GPUs.
  • H3 draft then final, same audio: Renders a cheap 4-step MiniMax H3 draft at 896x512, then the 20-step full-resolution final with the draft's sampled audio transplanted unchanged (Molinist MiniMaxH3TransplantAudioLatent), so the voice, words and timing…
  • H3 draft then final, same audio (first frame): The draft-then-final H3 job from a keyframe: both takes start on the same first frame, so the final keeps the draft's staging as well as its exact audio, re-rendered at full resolution.
  • H3 Extend: Continues an H3 take by one more window with H3 Motion Context: it reads only the clip's last 2 s, whose frames and sound carry the motion across the join, and returns only the new frames and sound, to place right…
  • H3 face refine (tracked crop): Repairs a small, smeared face in an H3 take: the face is tracked on every frame, cropped to fill a 768 px canvas, re-rendered by MiniMax H3 Ref2VA against the identity plate with the source speech held, and pasted…
  • H3 one-pass long take (40 s, no cuts): Animates a reviewed still as one unbroken take of up to 40 s with self-hosted MiniMax H3: one 957-frame window at 736x416 in eight steps, so there are no joins to hide, at about a third of the cost of chaining short…
  • H3 one-pass long speaking take (40 s): A reviewed still says a supplied voice line of up to 40 s in one unbroken take, rendered by self-hosted MiniMax H3 at 736x416 in eight steps with the line as the take's audio guide, so every word is the line's and…
  • Person swap: Keep the motion and camera of one 2–15s clip and replace the person with an identity still. Official MiniMax H3 reference-to-video at stereo 768p. Discover Try waits until both attachments exist.
  • H3 re-camera (CrossView-Warp): Re-shoots an existing clip from a new camera angle or along a keyframed camera move: MoGe-2 geometry and the CrossViewWarp node reproject the source, and the CrossView-Warp LoRA on MiniMax H3 Ref2VA fills what the…
  • H3 native 1088p from a still (turbo): Animates a reviewed still with self-hosted MiniMax H3 at native 1920x1088 (or 1088x1920) in eight steps with the Comfy-Org 768p turbo LoRA: about 2.3x faster and 60% cheaper than a 20-step native 1088p take, with most…
  • H3 native 1088p speaking take (turbo): A reviewed still says a supplied voice line, rendered by self-hosted MiniMax H3 at native 1920x1088 (or 1088x1920) in eight steps with the line as the take's audio guide, so a talking face reaches a 1080p delivery…
  • H3 native 1088p from text (turbo): Text-to-video with self-hosted MiniMax H3 at native 1920x1088 (or 1088x1920) in eight steps with the Comfy-Org 768p turbo LoRA, with stereo sound, for a 1080p delivery without an upscale.
  • H3 Long Take: Next Window (Latent Carry): Continues a long H3 take from the previous job's saved latent with H3 Motion Context: no decode between windows, and the new frames join the last window frame-exact.
  • H3 Long Take: Anchored Window (Latent Carry): Continues a long H3 take from the previous job's saved latent toward a closing keyframe you supply: H3 Motion Context carries the motion in, and the window ends on your frame.
  • H3 Long Take: First Window (Latent Carry): Starts a long H3 take from a keyframe, one window per job: returns the window and its sampler latent so the next job continues the shot without decoding.
  • Hosted FLUX.2-dev image editing (B200×4): FLUX.2-dev image editing on 4 B200 GPUs. Uses 20-step FP8 inference.
  • Hosted MiniMax H3 image-to-video (engine): Hosted MiniMax H3 image-to-video (engine). Authored Denoise Engine node; standard rail.
  • Hosted MiniMax H3 Turbo image-to-video (B200×4): MiniMax H3 Turbo image-to-video on 4 B200 GPUs. Uses the five-step Turbo preset.
  • Hosted MiniMax H3 reference-to-video (engine): Hosted MiniMax H3 reference-to-video (engine). Authored Denoise Engine node; standard rail.
  • Hosted MiniMax H3 text-to-video (engine): Hosted MiniMax H3 text-to-video (engine). Authored Denoise Engine node; standard rail.
  • Hosted MiniMax H3 Turbo text-to-video (B200×4): MiniMax H3 Turbo text-to-video on 4 B200 GPUs. Uses the five-step Turbo preset.
  • Hosted FLUX.2 Klein image editing (B200×1): FLUX.2 Klein image editing on 1 B200 GPU. Uses four-step inference.
  • MiniMax H3 experimental 4K text-to-video (H3-Base, not MiniMax 2K): Experimental H3-Base 4K text-to-video on our GPUs. Not MiniMax 2K and not hosted Hailuo.
  • MiniMax H3 self-hosted continue: Tail-guided MiniMax H3 continue on our GPUs. Pins the prior 22-frame tail through Add Guide.
  • MiniMax H3 hybrid continue: Iterate-only first-frame lock continue (no AddGuide). Concat + h3-join onto the prior file.
  • MiniMax H3 self-hosted face refine: Surgical H3 face restage on our GPUs. Requires likeness consent.
  • MiniMax H3 Draft 480p first-last-frame: Draft 480p MiniMax H3 first-last-frame interpolation on our GPUs. Standard is self-host stereo 768p; Hailuo hosted 2K stays a separate moderated option.
  • MiniMax H3 self-host stereo 1440p first-last-frame: Self-host stereo 1440p (2560×1440, 73 f) MiniMax H3 first-last-frame interpolation on our GPUs. Hailuo hosted 2K stays a separate moderated option.
  • MiniMax H3 experimental 4K first-last-frame (H3-Base, not MiniMax 2K): Experimental H3-Base 4K first-last-frame on our GPUs. Not MiniMax 2K and not hosted Hailuo.
  • MiniMax H3 self-host stereo 768p first-last-frame: Self-host stereo 768p MiniMax H3 first-last-frame interpolation on our GPUs. Hailuo hosted 2K stays a separate moderated option.
  • MiniMax H3 Lightning 8-step 768p first-last-frame: Official Comfy Lightning 8-step on 768 FLF. Composer opt-in only — omitted from Discover rails, Autopilot, and Catalog Guide.
  • MiniMax H3 8-step first-last-frame: Lab-only Acc-LoRA FL2VA graph. Stock LoraLoaderModelOnly loads the Alibaba PAI 8-step file but does not cut wall-clock versus 20-step on Comfy — not a cheap Discover SKU.
  • MiniMax H3 Draft 480p image-to-video: Draft 480p MiniMax H3 image-to-video on our GPUs. Standard is self-host stereo 768p; Hailuo hosted 2K stays a separate moderated option.
  • MiniMax H3 self-host stereo 1440p image-to-video: Self-host stereo 1440p (2560×1440, 73 f) MiniMax H3 image-to-video on our GPUs. Hailuo hosted 2K stays a separate moderated option.
  • MiniMax H3 experimental 4K image-to-video (H3-Base, not MiniMax 2K): Experimental H3-Base 4K image-to-video on our GPUs. Not MiniMax 2K and not hosted Hailuo.
  • MiniMax H3 self-host stereo 768p image-to-video: Self-host stereo 768p MiniMax H3 image-to-video on our GPUs. Hailuo hosted 2K stays a separate moderated option.
  • MiniMax H3 768 draft + LTX-2.5 refine (experimental): Experimental 4-step H3 draft then 3-step LTX-2.5 refine. Not an acceleration profile. Never Auto.
  • MiniMax H3 Turbo 8-step 768p image-to-video (experimental): Experimental 8-step Turbo option for self-host 768p MiniMax H3 image-to-video. Composer opt-in only once proven — never Auto, never Discover.
  • MiniMax H3 Draft 480p last-frame: Draft 480p MiniMax H3 last-frame lock. Standard is self-host stereo 768p; Hailuo hosted 2K stays a separate moderated option.
  • MiniMax H3 self-host stereo 768p last-frame: Self-host stereo 768p MiniMax H3 last-frame lock. Hailuo hosted 2K stays a separate moderated option.
  • MiniMax H3 detail pass: H3 detail pass (2× then back to source size). Re-encode the clip through the H3 video encoder. This is not 2K and not Hailuo regenerate.
  • MiniMax H3 motion-context chain: Productions starter: two-clip native H3 graph that pins previous-clip latent into the next clip and writes clip_00002.safetensors. Not a Discover SKU.
  • MiniMax H3 Draft 480p reference-to-video: Draft 480p MiniMax H3 reference-to-video on our GPUs. Standard is self-host stereo 768p; Hailuo hosted 2K stays a separate moderated option.
  • MiniMax H3 self-host stereo 1440p reference-to-video: Self-host stereo 1440p (2560×1440, 73 f) MiniMax H3 reference-to-video on our GPUs. Hailuo hosted 2K stays a separate moderated option.
  • MiniMax H3 experimental 4K reference-to-video (H3-Base, not MiniMax 2K): Experimental H3-Base 4K reference-to-video on our GPUs. Not MiniMax 2K and not hosted Hailuo.
  • MiniMax H3 self-host stereo 768p reference-to-video: Self-host stereo 768p MiniMax H3 reference-to-video on our GPUs. Hailuo hosted 2K stays a separate moderated option.
  • MiniMax H3 8-step reference-to-video: Lab-only Acc-LoRA Ref2VA graph. Stock LoraLoaderModelOnly loads the Alibaba PAI 8-step file but does not cut wall-clock versus 20-step on Comfy — not a cheap Discover SKU.
  • MiniMax H3 RefMod reference-to-video: Lab-only native H3 reference-to-video graph that injects compressed reference mods instead of official still slots. Not a Discover SKU.
  • MiniMax H3 Draft 480p text-to-video: Draft 480p MiniMax H3 on our GPUs. Standard is self-host stereo 768p; Hailuo hosted 2K stays a separate moderated option.
  • MiniMax H3 self-host stereo 1440p text-to-video: Self-host stereo 1440p (2560×1440, 73 f) MiniMax H3 on our GPUs. Hailuo hosted 2K stays a separate moderated option.
  • MiniMax H3 self-host stereo 768p text-to-video: Self-host stereo 768p MiniMax H3 on our GPUs. Hailuo hosted 2K stays a separate moderated option.
  • MiniMax H3 Draft 480p video edit: Draft 480p MiniMax H3 video-to-video edit. Standard is self-host stereo 768p; Hailuo hosted 2K stays a separate moderated option.
  • MiniMax H3 self-host stereo 768p video edit: Self-host stereo 768p MiniMax H3 video-to-video edit. Hailuo hosted 2K stays a separate moderated option.
  • H3 4K video: text, image, first/last frames and references: Explicit native H3 graph revisions generate on the required 32-pixel grid and crop to exact 3840x2160 delivery while preserving original model and media paths.
  • H3 long-take continuous shot and audiovisual continuation graphs: Long continuous shots, including 60-second plans: bounded H3 windows with direct or saved motion-context AV latents, encoded AV/frame guides, last-frame hybrid continuation and exact overlap assembly. Includes an…
  • Extend: More seconds after the last frame, same scene.
  • Reference to video: A clip with your person, product or place in it.
  • Text to video: A clip from a sentence.

Setup and requirements

  • Name the option you want in your brief; the agent reads its method or configuration (Source entry points) and applies it.

Source licenses and model terms continue to apply. Check runtime availability before generation.

Download source · Learn how to reuse resources

Related resources