Files
DumasNodes/docs/TURBO_SAMPLE.md
T
chris.dumas bf8f76505f Squashed 'vendor/dumas_h3_continuum/' content from commit d07ed31
git-subtree-dir: vendor/dumas_h3_continuum
git-subtree-split: d07ed31b442d283d6a1740f875a843b9feda4c91
2026-10-20 13:43:38 +00:00

3.1 KiB

V3.9 Turbo Sample

A separate, user-supplied example of Continuum prompt formatting, not a new model or a verified quality preset. Available in Templates as MiniMax_H3_Continuum_V39_Turbo_Sample (no parenthesized suffix).

  • Prompt Format: Auto; 2 chunks x 8 seconds; Full Run; Progress OFF.
  • The multiline prompt demonstrates [Chunk 1], [Chunk 2], first-image alignment, continuous dancing with a fixed background/camera, and an off-screen male Thai rap. <d>[Thai] ... </d> is H3 lyric syntax, not part of the template name.
  • Saved model: minimax/DasiwaMinimaxH3_dasiwaHybridV4_int8.safetensors. Obtain it from the author's model page or select a compatible installed H3 model. No model weights are bundled.
  • One LoRA is enabled: minimax/minimax_h3_fl2v_turbo_4step_v1.2_768p_comfyui_bf16.safetensors, strength 1, from LightX2V. The supplied graph uses Euler / simple / 6 steps / Draft 0.30 MP. These are preserved sample values, not a validated 4-step/768p configuration; check the selected model and LoRA instructions before generation.
  • Video VAE: minimax_h3_video_vae_int8_convrot.safetensors; audio VAE: minimax_h3_audio_vae_fp32.safetensors; text encoder: qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors, type minimax. See model download links and folders.
  • Select your own First Image (example.png is only a placeholder). The disabled optional audio/video inputs retain sample filenames; select real files before enabling them. Reselect model/LoRA paths if you do not use a minimax subfolder.
  • Requires KJNodes (SageAttention and Model Preview Override), rgthree, and Spectrum node definitions. Spectrum is OFF. Preview Override has video/audio VAE connected; this is outside the audited v3.9.8 Preview Phase A Take-reuse profile.
  • Thai lyric accuracy, unchanged background, identity preservation and subjective video/audio quality are not established by adding this sample. Thai is outside H3's explicitly listed stable dialogue languages. No new GPU test is claimed.
  • Existing standard workflows and saved Runs/Takes are unchanged. Outputs use the separate video/MiniMax_H3_Continuum_V39_Turbo_Sample prefix. GitHub main includes this sample; the already-published v3.9.8 release/Registry package is not replaced.

日本語

標準Workflowとは別のTurboサンプルです。Auto・2チャンク×8秒で、改行した 映像・環境音・音楽の記述、入力画像の指定、固定背景、画面外の男性タイ語ラップを 示します。画像・モデル・LoRAは各自で導入して選択してください。 4-step/768p LoRAに対して6 steps・Draft解像度という添付の設定を保持しています。 推奨設定や画質の合格を意味しません。タイ語の発音、人物・背景保持も未保証です。 KJ Preview OverrideのVAE接続はv3.9.8の限定Take再利用検証の範囲外です。 今回はサンプル追加とCPU検証のみで、実環境反映・新規GPU試験・Release更新は行いません。