Skip to content

Self-hosted, Self-Improving, Federated Everything App — dev management, agent orchestration, persistent mind, digital twin, second brain, and creative pipeline from a single dashboard over Tailscale to port the user into the OS

3.9 KB
Raw Blame
// Advice uses runtime capabilities and declared Finish edges, never a guessed
// cross-model seed relationship. Changing size does not start a generation.
export default function ModelWorkflowHelp({ model, mode, models = [], onResolutionChange }) {
  if (!model) return null;
  const fastH3 = model.runtime === 'fastvideo' && model.fastvideoFamily === 'fasth3';
  const fastMetal = model.runtime === 'fastvideo' && !fastH3;
  const finish = models.find((entry) => entry.id === model.finishModelId);

  return (
    <>
      {mode === 'image' && model.runtime === 'minimax_h3' && (
        <p className="mt-2 text-xs text-gray-400" role="note">
          MiniMax H3 can replace the source scene when the prompt describes a different environment.
          Start with a motion-only prompt and a model-native resolution, then inspect the whole clip.
          Image conditioning does not guarantee scene continuity.
        </p>
      )}
      {mode === 'image' && models.some((entry) => entry.runtime === 'fastvideo' && !entry.supportedModes?.includes('image')) && (
        <p className="mt-2 text-xs text-gray-400" role="note">
          Image mode lists models with image conditioning in their installed runtime. The shipped FastMetal
          and FastH3 MLX runners support text-to-video only, so they are available in Text mode.
          For image-to-video, choose a compatible LTX, Wan, or MiniMax H3 model listed above.
        </p>
      )}
    <details className="mt-2 rounded-lg border border-port-border px-3 py-2 text-xs text-gray-400">
      <summary className="cursor-pointer text-gray-300">Choosing a model: preview → final</summary>
      <div className="mt-2 space-y-2 leading-relaxed">
        <p>
          Start with a small resolution and the shortest available duration to check composition and motion.
          Smaller renders reduce generation work, but model loading and first-use conversion still take time.
          Keep a distilled model’s trained step count; fewer steps can spoil the result.
        </p>
        {fastH3 ? (
          <>
            <p>
              {model.fastvideoVsa
                ? 'FastH3 V2 is an eight-step quality option with video and audio, not a real-time preview model.'
                : 'FastH3 Preview uses four steps for video and audio; it still loads a large model.'}
              {' '}Use the 832×480 preset and shortest duration for an initial quality check.
              For a cheaper wiring test, try 512×288; this smaller canvas is experimental and may reduce quality.
              INT6 uses smaller transformer weights than INT8; it is not a guaranteed speed increase.
            </p>
            <button type="button" onClick={() => onResolutionChange(512, 288)} className="text-port-accent underline">
              Use 512×288 test size
            </button>
          </>
        ) : fastMetal ? (
          <p>FastMetal is a draft option: 1.3B is the smallest model, while 5B offers a larger three-step model. Try a small preview resolution before a larger render.</p>
        ) : (
          <p>Where offered, use the Fast speed profile or a Lightning model for drafts. Choose Quality for the final render. Original MiniMax H3 is a slow quality reference; smaller quantization mainly reduces weight size.</p>
        )}
        {finish && (
          <p>This model has a declared delivery pair: <strong className="text-gray-300">{finish.name}</strong>. Eligible text-to-video gallery clips offer Finish to re-render with that model.</p>
        )}
        <p>
          To enlarge a clip you like, use its gallery Upscale action: resize preserves the content;
          the LTX-2.5 generative option can add detail and change it, and requires its model pack.
          Switching from FastMetal or FastH3 Preview to FastH3 V2 generates a new shot — the same prompt
          and seed do not preserve the original composition across models. Use Finish only where offered.
        </p>
      </div>
    </details>
    </>
  );
}