Marcel Hofer
e0985b4f7b
Refactor the monolithic QwenVL implementation into focused core modules for model access, Hugging Face runtime and configuration, prompt state, prompt contracts, and media helpers while preserving the public compatibility facade. Add duration-aware LTX 2.3 and MiniMax H3 prompting, including effective-duration cache identity, MiniMax frame-grid normalization, dedicated T2VA/I2VA/FL2VA/L2VA/reference contracts, and best-effort output normalization. Improve streamed output handling by separating reasoning from final content, rejecting unfinished thinking traces, sanitizing terminal-only punctuation for legacy Windows code pages, and retaining original Unicode in responses and traces. Expand preset recommendations, ComfyUI tooltips, system prompts, and documentation for video token budgets, thinking settings, reference roles, dialogue fidelity, and duration behavior. Keep HF and GGUF consumers on canonical shared imports, preserve mutable export identities across reloads, and extend regression coverage for compatibility, downloads, caching, prompting, streaming, and node registration.
3.3 KiB
Video preset settings
ThinkingLLM shows two independent recommendation blocks in the read-only recommended_settings info box:
- The selected model family's sampler guidance.
- The selected video preset's output budget and thinking-mode guidance.
The info box never changes saved widget values. Preset guidance is centrally defined in AILab_System_Prompts.json and mirrored to the web payload used by ComfyUI.
Preset recommendations
| Preset family | enable_thinking |
max_tokens |
|---|---|---|
| LTX 2.3 T2V, I2V, First/Last, Reference | false |
512 |
| MiniMax H3 T2VA, I2VA, FL2VA, L2VA | false |
2048 |
| MiniMax H3 Full Reference | false |
3072 |
| Wan 2.2 Scene 3s/5s, T2V and I2V | false |
512 |
| Wan 2.2 Timeline 3s/5s, T2V and I2V | false |
768 |
| Wan 2.2 Scene 20s, T2V and I2V | false |
1536 |
| Wan 2.2 Timeline 20s, T2V and I2V | false |
2048 |
Thinking is disabled in these recommendations because the selected system prompts already define the planning and output format. Direct generation is faster and less likely to spend the output budget on a reasoning trace or deviate from the requested final structure.
The token budgets are upper bounds for the expected final format rather than settings from the downstream video generator. Increase the budget only when a useful final prompt is genuinely truncated.
Sampling settings
Temperature, top_p, top_k, and repetition penalties remain model- and input-mode recommendations:
- Qwen3/Qwen3.5 vision input generally uses
temperature=0.7,top_p=0.8,top_k=20, andrepetition_penalty=1.0. - Text-only input follows the selected model card. Qwen3.5 uses
temperature=1.0,top_p=1.0,top_k=20, andrepetition_penalty=1.0; Qwen3-VL Instruct usestemperature=1.0,top_p=1.0,top_k=40, andrepetition_penalty=1.0. - Widgets not exposed by the selected node are omitted from the info box summary.
For a connected reference video, begin around two sampled frames per source second and stay within the node's 64-frame limit. Image-only presets do not need a frame_count recommendation.
Format rationale
- LTX requests one literal chronological paragraph under 200 words, so 512 output tokens provide sufficient headroom.
- MiniMax Base uses three structured fields; Full Reference uses six sections and needs a larger budget.
- Wan single-scene prompts fit within 512 tokens. Timeline formats and four-part 20-second sequences need progressively larger budgets to avoid truncation.
Runtime-setting instructions such as max_tokens, context length, batch size, and temperature are kept out of the system prompts themselves. They belong in the info box and must not become part of the generated video prompt.