MiniMax H3 case archive — page 51 of 51

Browse source-attributed MiniMax H3 video cases 2401–2448 of 2448.

How to compare these examples

Choose a mode by your inputs: T2VA starts with text, FL2VA uses a first or last frame, and Ref2VA uses reference assets. Narrow the library by scene and style, then open an example to examine its output and source notes.

Compare subject consistency across cuts, camera continuity, and the timing of dialogue or environmental sounds against picture events. Resolution and duration describe the output; speed, VRAM needs, and reliability require checks on your own hardware.

Case summaries are our viewing notes. Videos, images, and complete prompts shown verbatim belong to their creators. When a full prompt is unpublished, consult the available assets and workflow; the original source determines the reproduction conditions.

  1. 2401 · Unknown · 2026-08-10Returning to a Simpler Workflow after Acceleration TestsAfter trying several MiniMax H3 workflows and acceleration methods, the creator returns to a simpler setup, waits for official 2K upscaling and speed support, and continues prompt research. The uploaded test is a 10.125-second portrait-format clip.@ai_haruki3
  2. 2402 · FL2VA · 2026-08-10Three-Shot FL2VA Chain on an RTX 5060 TiA home local-generation test using MiniMax H3 FL2VA pruned NVFP4, ComfyUI 0.30.0, SageAttention, and Qwen3VL-32B NVFP4 AWQ on an RTX 5060 Ti 16GB. Twenty-five steps chain three L2VA-to-I2VA-to-I2VA shots into a 15.532-second clip at 832 by 1248.カブ
  3. 2403 · Unknown · 2026-08-10Rino's TerritoryA 15.083-second square MiniMax H3 clip titled Rino's Territory. The post provides only the title and model tag, so no character setup, input method, or prompt is inferred.dimesfromthevoid
  4. 2404 · Unknown · 2026-08-10Street Portrait Consistency ComparisonThe creator compares this MiniMax H3 example with a similar video made locally with Wan 2.2 about a year earlier and reports a substantial gain in consistency and expressiveness. The clip shows a character close-up on a Japanese street. No prompt is published.@kiyoshi_shin
  5. 2405 · Unknown · 2026-08-10Local 480p Generation Benchmark on an RTX 5090The creator reports a local MiniMax H3 benchmark on an RTX 5090: a 15.083-second 480p clip generated in one minute and 27 seconds. The video depicts a warmly lit futuristic cabin interior. No prompt is published.STRIMANDO.
  6. 2406 · Unknown · 2026-08-10Surreal H3 Variant Built on KreaThe creator built a MiniMax H3 version of the same concept on Krea and quotes an earlier Seedance 2.5 version for comparison. The vertical clip features a giant eye surrounded by dense cone-shaped one-eyed figures. No prompt is published.vkuoo
  7. 2407 · Ref2VA · 2026-08-10Suno Beat-Synced Character Montage Across Five EnvironmentsA character sheet and a 15-second Suno track guide 30 shots across five environments, with cuts tied to strong beats, half-beats, and musical phrases. The post publishes the full prompt for identity consistency, shot structure, audio synchronization, and negative constraints.Kōda
  8. 2408 · T2VA · 2026-08-10Handheld Space-Cruise POV at a Black HoleA tourist records a handheld view from a space-cruise cabin as a giant black hole fills the window. They switch off the light to improve the view and speak a short line. The post publishes the complete prompt verbatim.Ivan Fioravanti ᯅ
  9. 2409 · Unknown · 2026-08-10Thirty-Second Continuous Speech and Lip SyncA full 30-second portrait dialogue test. The creator explicitly describes it as one generation rather than a stitched sequence and reports continuous speech without noticeable sync drift. The post does not publish a prompt.Rich · Atom Tan Studio
  10. 2410 · Ref2VA · 2026-08-10Ref2VA Frame-Infill Cameo InsertionA new Ref2VA node for ComfyUI-MiniMaxH3FrameInfill demonstrates regenerating only seconds 3.0–4.5 of a dance clip to insert a cameo, plus timed entrances, exits, and a high-five for two characters. The creator explicitly labels the feature experimental.@redpolo999
  11. 2411 · Ref2VA · 2026-08-10Beat-Synced Spaghetti-Western Title SequenceA character reference and audio track drive a 10-second spaghetti-western title sequence whose hard cuts, freeze frames, silhouettes, stencil typography, and distressed-film effects land on the beat. The full prompt is published in the post.PhotoGPT
  12. 2412 · T2VA · 2026-08-10Local No-Reference H3 Test on an RTX 5060 TiCardoso runs H3 locally on an RTX 5060 Ti 16GB and publishes a 14-second low-resolution test. The post explicitly states that no external reference was used and highlights the model-generated voices, intonation, and expressions.cardoso
  13. 2413 · T2VA · 2026-08-10Four-Minute Local T2V Sitcom TestA locally produced T2V sitcom-coherence test with a total X upload length of four minutes and 34 seconds. The creator says new nodes are improving long-sequence consistency. The upload length is not presented as a single H3 generation.alawlette
  14. 2414 · FL2VA · 2026-08-10One-Image Bobblehead in Photoreal and Cartoon PassesOne desk photo of a bobblehead becomes two five-second passes, one photoreal and one 2D cartoon. The creator says picture and audio were generated together without added foley or text-to-speech, entirely on an RTX 5090.Mike Frost
  15. 2415 · Unknown · 2026-08-10Storyboard-and-Prompt Text Rendering TestA 15-second H3 test combines a storyboard with prompting across widescreen scenes and on-screen text. The creator specifically notes that the text follows prompt instructions, but does not publish the prompt verbatim.PeL：I&P
  16. 2416 · Unknown · 2026-08-10Local 1080p Generation Benchmark on an RTX 5090A public local-performance record for generating a 10-second 1080p clip on an RTX 5090, with a reported runtime of 15 minutes and 23 seconds. An adjacent post reports one minute and 27 seconds for a 15-second 480p version, illustrating the resolution cost.STRIMANDO.
  17. 2417 · Ref2VA · 2026-08-10Cloud-Head Subway TransformationA locked-off, symmetrical subway shot glitches the passengers' heads between human faces and distinct weather clouds, then converges on the reference composition at eight seconds. The original post also publishes the complete visual and sound prompt.AION
  18. 2418 · Ref2VA · 2026-08-10Rainy Bar Continuous-Tracking ComparisonThe same rainy-street concept is tested with Seedance 2.5 and ComfyUI MiniMax H3. The H3 version outputs a 15-second, 1376×768 continuous tracking shot with timed constraints for moving from a frontal approach to a rear follow. The post publishes both complete prompts.FoldingEnd A.I.
  19. 2419 · Unknown · 2026-08-10Single-Image Dialogue and Prosody TestUsing only one character reference image, the creator generates a 15-second indoor two-person dialogue scene and specifically praises H3's handling of voice, intonation, and prosody. The post does not publish a prompt.cardoso
  20. 2420 · Unknown · 2026-08-10Locally Generated Isekai Anime SequenceAn approximately 29-second isekai slice-of-life anime short explicitly labeled as a local H3 generation. In a reply, the creator says it began as two 15-second sections, with the forest portion edited from two or three clips.Mal
  21. 2421 · Unknown · 2026-08-10Local Character Dance on an RTX 3080A 10-second portrait character dance generated locally on an RTX 3080 and Core i7-13700K. The creator says they borrowed another creator's dance prompt and liked the musical pauses and ending, but this post does not reproduce the prompt verbatim.Toshi@ニャルオAI
  22. 2422 · Unknown · 2026-08-10Lip-Sync Breakdown After Ten SecondsA 15-second stage-singing lip-sync test. The creator explicitly says the lyrics break down after roughly ten seconds, so the audio was replaced, while the replacement still appears reasonably aligned with the mouth movement.Alone1M
  23. 2423 · Unknown · 2026-08-10Local Magical-Girl Transformation on a 4070 TiA 10-second magical-girl transformation sequence that uses rapid motion to mask local artifacts. The creator reports that an RTX 4070 Ti 12GB takes roughly nine minutes to generate about ten seconds locally.@NOTorJB
  24. 2424 · Unknown · 2026-08-09Same Prompt: H3 vs Seedance 2YukYuk generated separate 15-second videos with MiniMax H3 and Seedance 2 using the same prompt to compare their outputs. The original post does not disclose the prompt text.YukYuk!
  25. 2425 · Unknown · 2026-08-09Local ComfyUI Generation on RTX 3090The creator generated a 15-second H3 video locally in ComfyUI on an RTX 3090. The published output is 864×480, took approximately 22 minutes to render, and was presented as a positive test of prompt adherence.Melvin Vivas
  26. 2426 · T2VA · 2026-08-09One-Line Prompt: Eating RamenThis local Turbo LoRA experiment uses only the creator’s published one-line prompt, “Eating Ramen,” to test the character’s hands, chopsticks, and eating motion.CALMDUST aka ちゃんこあ
  27. 2427 · Unknown · 2026-08-09Local Audio-Video Generation on Mac StudioThe creator states that both the 37-second video and its audio were generated locally with MiniMax H3 in a single Claude Code run. The disclosed hardware is a Mac Studio with an M3 Ultra and 256 GB of memory.@rsensui
  28. 2428 · T2VA · 2026-08-09Hand-Painted Blizzard SurvivalThe creator publishes the full prompt for a frame-by-frame oil-on-paper animation with a 12 fps stop-motion cadence, cold moonlight, and a violent blizzard, depicting a mother and child struggling to survive extreme weather.sada
  29. 2429 · FL2VA · 2026-08-09Krea 2 Image to MiniMax H3Uses an image generated with Krea 2 as the input for H3, creating an image-to-video workflow suitable for overnight batch rendering. The original post does not disclose the prompt.Tom Maiaroto
  30. 2430 · FL2VA · 2026-08-09Interactive Deep-Sea Bestiary UIUses the same full prompt to compare H3 with Seedance 2.5: a first-frame-locked deep-sea encyclopedia interface in which only the cursor, main title, and creature in the left panel may change, with precisely timed audio and extensive negative constraints.Jin.B
  31. 2431 · Unknown · 2026-08-09Long-Form MiniMax H3 Music VideoA MiniMax H3 music-video experiment lasting approximately 2 minutes and 17 seconds. The original post identifies the model but does not disclose the prompt or input assets.Jinny Lee
  32. 2432 · Unknown · 2026-08-09Local 12-Step Fluid and Fur TestsThree local ComfyUI tests, each lasting five seconds, use speed-first 12-step settings to examine H3’s handling of fluids and fur. The creator notes that both remain challenging at low step counts.Jun Ozaki
  33. 2433 · Ref2VA · 2026-08-09Ref2VA Character Sheet Consistency TestUses a character sheet generated with ChatGPT as the reference input for an H3 Ref2VA consistency test. Published parameters include an RTX 4070, 576×832 resolution, 24 fps, a five-second duration, and an approximately 225-second generation time.Kamimoto
  34. 2434 · Unknown · 2026-08-092K Flight Scene Upscaled to 4KThe creator first generates a 2K flight scene with MiniMax H3 and then upscales it to 4K with Astra, demonstrating a workflow that connects generation with post-production super-resolution.AION
  35. 2435 · FL2VA · 2026-08-09I2V Turbo LoRA TestA 10-second MiniMax H3 image-to-video Turbo LoRA test created in ComfyUI. The original post does not disclose the reference-image details or prompt.toyxyz
  36. 2436 · Ref2VA · 2026-08-09Nine-Reference AI Film TrailerH3 Omni processes nine reference images simultaneously in a 15-second AI film trailer while also testing fine-detail text rendering. The creator only discloses that the prompt contains approximately 3,943 Chinese characters, without publishing its contents.Cia0
  37. 2437 · Unknown · 2026-08-09The White Camellia: Six-Shot Local FilmThe White Camellia comprises six 15-second H3 shots and combines Qwen Image 3 Pro with ElevenLabs. It was produced locally on an RTX Pro 6000 with 96 GB of memory.Nedelcho Sitnov
  38. 2438 · FL2VA · 2026-08-09Larry Turbo and H3 Sage Acceleration TestA 1 MP, 10-second I2V acceleration test on an RTX 4070 took 11 minutes and 12 seconds, followed by upscaling and frame interpolation. The creator also documents blocky facial artifacts and weak speech generation.AIconia
  39. 2439 · Unknown · 2026-08-09Five-Scene Rainy Tokyo Music VideoMotion Context chains five scenes together, uses H3-generated vocals as the music, and adds beat-synchronized lyrics with HyperFrames to create an approximately 29-second music video set in rainy Tokyo.Maki
  40. 2440 · T2VA · 2026-08-06The Diner That RewindsInside a 1950s diner, coffee, dishes, and food freeze in mid-air after a collision. The camera circles the scene before the entire world rewinds to its starting state. The original post includes a segmented camera-and-action prompt, although the displayed video lasts 15 seconds while the prompt timeline spans 30 seconds.ICREAT AI
  41. 2441 · FL2VA · 2026-08-06Textbook Scene ReimaginedThe creator used an image to establish the visual style and a text prompt to recreate a primary-school textbook scene, structuring the 15-second video around the emotional arc of Beethoven’s Moonlight Sonata—from calm to turbulent and back to calm.@Yangtze_Seventh
  42. 2442 · Unknown · 2026-08-06Composition — Layered Music VideoA long-form music video built from Suno music, Midjourney visuals, and MiniMax H3 video generation. The creator highlights H3’s object layering and prompt adherence, but does not disclose the exact generation mode or shot-level prompts.tapehead.lab
  43. 2443 · T2VA · 2026-08-06One Prompt, Two CommercialsUses the same prompt and settings in Pollo AI to compare commercial outputs from MiniMax H3 and Seedance 2.0. The creator describes H3 as more natural and polished, while the other model places greater emphasis on cinematic visual impact.George
  44. 2444 · Unknown · 2026-08-06Localized UGC Product FilmAn AI UGC product-video workflow tailored to different countries and content platforms. The creator states that the commercial version mainly combines HyperFrames with MiniMax H3, although the division of responsibilities between them remains undocumented.Tuola Ge
  45. 2445 · Ref2VA · 2026-08-06Person Replacement TestA ComfyUI-based person-replacement video experiment. The original post explicitly identifies MiniMax H3 but provides no prompt, reference inputs, or node configuration, and includes a visual reference to a copyrighted character.Nobody
  46. 2446 · T2VA · 2026-08-05After the Fleet JumpsMoves from a medium-wide view of a starship bridge to a close-up of the captain, using native stereo audio to synchronize the fleet jump, spatial shock, and the suddenly quiet aftermath.MiniMaxAI
  47. 2447 · FL2VA · 2026-08-05Ramen Rack FocusUses a ramen image as the opening-frame constraint, then gradually racks focus so the foreground steam softens while the family in the background comes into view, testing depth of field and ensemble motion.MiniMaxAI
  48. 2448 · Ref2VA · 2026-08-05Follow the WindReferences a source video, its original background music, and an additional voice timbre to add English dialogue while preserving character identity, wardrobe, framing, and setting.MiniMaxAI
← Previous