ComfyUI Node

T8 Creator Style Config(导演风格参考)

How to say Fincher without typing 'Fincher'

By T8mars·Created 3 months ago·Updated 5 days ago· 409
T8 Creator Style Config(导演风格参考)
    • creator_style_config
    ◄mode关闭 / Off►
    ◄creator未选择 / None►

    A name is not a shot list

    Here's the trap this node dodges. Type "shot like Fincher" into a video prompt and you've said almost nothing a model can act on - you handed it a surname and hoped. The usual LLM-enhancer failure is the mirror image: you ask for detail, get detail you never specified, and your one-line idea comes back with a plot you didn't write.

    T8 Creator Style Config goes the other way. You pick one of 200 catalog entries by name, and the node hands your H3 or Seedance 2.0 enhancer a concrete list of presentation methods attached to that entry - framing, camera, staging, rhythm, how information is ordered, editing - instead of the name. The author's one-liner is "presentation reference, not a screenplay generator," and the injected instruction block spends most of its words forbidding things: no borrowed plots, no inferred traits, no props, cuts, reveals or reactions that weren't already in your prompt.

    It shares its 200-entry catalog with T8 Color Style Config but stays independent: choosing a director doesn't choose their palette. That's the right call - half of "why does everything look the same" traces back to look-books that bundle camera grammar, color and lighting into one blob.

    How it works

    Same mechanism as the color node: a typed payload carrying identity, a resource revision and a SHA-256, not prose. The enhancer re-derives the canonical version from the shipped catalog, rejects anything that doesn't match, loads resources/visual_styles/creators/<id>.json, verifies the hash, and appends a system block to the request it was already making.

    That resource holds one to eight methods, each tagged with a dimension (framing, camera, staging, rhythm, information, editing) and a requires_cuts boolean. The block lists them by dimension and fences them in: use existing space and authorized actions only, let framing guide attention without adding story facts, props, cuts or reveals, and keep who can perceive each event - a presentation suggestion can't leak a later secret or change its cause.

    There's a detail in the source that reads like a bug when you hit it: if the enhancer run is pinned to exactly one shot, methods marked requires_cuts are filtered out before the block is built, so a single-shot clip gets a shorter method list than the same entry at six shots. That's intentional - it won't feed the model a cut-based method it structurally can't perform. If a repair or retry fires, the style block is re-attached rather than dropped.

    The two inputs

    Mode defaults to 关闭 / Off, and Creator defaults to 未选择 / None. The list is the same 200 numbered entries the color node shows - 001 阿尔弗雷德·希区柯克 / Alfred Hitchcock onward - but here you're reading the presentation half, 400 methods spread across those 200 looks. There's no scope field, and that's not an omission: nothing in the creator entry can bleed into your color, so scope only exists on the color side.

    Output is one creator_style_config socket, going into the 导演风格(可选)/ Creator style input on the H3 or Seedance 2.0 enhancer. It's optional and lazy - unconnected or Off, the enhancer never evaluates it. Separate socket from color means director-only, color-only or both; the pack ships an example workflow for each wiring.

    What's actually in the catalog

    Be clear-eyed, because a "200 directors" pitch oversells itself. These aren't 200 film theories. The role field breaks down as 195 directors plus two cinematographers, one art director and two teams, and thinly sourced entries are flagged as generic shot templates rather than exclusive director theory.

    The browser is honest about the rest: previews are original vector diagrams explaining a composition or pacing idea, not film stills and not renders, and the author states that no verified model output exists for these entries yet. What you're buying is a shortcut past the blank page - a few concrete, non-obvious moves written in English your LLM can act on, attached to a look you can name.

    Installing it

    Both config nodes come with the same pack - install one, you have the other:

    cd ComfyUI/custom_nodes
    git clone https://github.com/T8mars/comfyui-minimax-h3-prompt-enhancer-T8.git
    

    Restart ComfyUI, Ctrl+F5 the browser. Or use ComfyUI Manager and search MiniMax H3 / Seedance 2.0 / Music 3 Prompt Enhancer (T8). No models, no API key at the config level, and no heavy Python deps - but the pack does require ComfyUI 0.33.0 or newer, since these nodes use the newer io.ComfyNode API and a custom socket type. Keep one copy of the pack in custom_nodes - duplicate copies are the classic route to mismatched resource hashes.

    Where people get burned

    Switching Mode to On with 未选择 / None still in the creator box fails at queue time with "Select a valid style before enabling it." Select first, enable second. If the browse dialog won't open, the pack's frontend and backend routes both need to be live - restart ComfyUI fully, not just the page. The browse button isn't serialized into your workflow either, so a shared workflow carries the saved selection, not any browsing state.

    Then the expectation traps. Don't wire this up expecting the prompt to say "Fincher" anywhere - it says "use existing frames or reflections to relate observer and target," and similar. If your prompt was empty of specifics, the modifier has nothing to organize. Don't fight the single-shot filtering when the method list looks short; check your shot count first. And don't treat it as a substitute for the pack's other config nodes: performance and character bibles own acting, combat owns fight staging, storyboard owns structure. This one owns presentation only, and says so in every request.

    Finally - experimental, in a pack with essentially no community footprint yet. Judge it with one A/B: same prompt, same seed, config Off then On. If the second version reads like it knows where the camera goes, keep it. If not, switch it off and save the tokens.

    CategoryT8/Utilities

    Inputs (2)

    NameTypeDefaultDescription
    modeCOMBO关闭 / Off3 options: 关闭 / Off, 开启 / On, 开启(H3强化试验) / On (H3 experimental)
    creatorCOMBO未选择 / None201 options: 未选择 / None, 001 阿尔弗雷德·希区柯克 / Alfred Hitchcock, 002 大卫·芬奇 / David Fincher, 003 迈克尔·哈内克 / Michael Haneke, 004 布莱恩·德·帕尔玛 / Brian De Palma, 005 丹尼斯·维伦纽瓦 / Denis Villeneuve, +195

    Outputs (1)

    NameTypeDescription
    creator_style_configT8_CREATOR_STYLE_CONFIG—