凤希AI - 场景加载器(简化版)

FxAiScreenLoadSimple (凤希AI - 场景加载器简化版) is the tiny scene loader at the heart of the fxai-toolkit long-video workflows. Give it a list of scenes and a loop index, and it hands you two things for the current scene: the 台词 (dialogue/script text, ready to become your prompt) and the 素材 (the reference images, ready to become your img2vid inputs). That's the whole job. If you're rendering video one scene per loop iteration - exactly how this pack generates "unlimited length" MiniMax H3 clips on low-VRAM cards - this is the node you want, and no more.

The pack itself is 凤希 (Fengxi) AI's "all-in-one" node collection: a Chinese-authored, low-spec-friendly long-video toolkit built around running MiniMax H3 locally, with support over a QQ group and Bilibili videos rather than an English issue tracker. It's MIT licensed and clearly still being worked on, but it's a product-first pack - the README is a sales manual, not an install guide. This node is one piece of a bigger machine, but a genuinely useful one.

How it works

The node takes a 场景数据 list and an integer 行索引, then does three small things:

1. It picks the row at 行索引, but modulo the list length. Index 7 on a 5-scene list quietly becomes scene 2. That's the "循环取值" the name hints at - a loop counter can run forever without ever throwing an out-of-range error. This is the whole point of the simplified version: the full FxAiScreenLoad raises IndexError on a bad index and also outputs frame counts, reference voices, and audio; this one strips all of that.
2. It builds the prompt as 通用提示词 + 台词 + 尾部通用提示词 - a header and footer you feed in once, wrapped around each scene's text. If a scene's dialogue is empty, it falls back to the first scene's text rather than emitting a blank prompt.
3. It loads the scene's images. The 素材 field is a comma-separated list of relative paths like avatar/007.png,scene/002.jpg, and here's the gotcha: those resolve under ComfyUI/fxai/image/, not ComfyUI/input/. Each image becomes an RGBA tensor and comes out the IMAGE port as a batch.

The pack's own FxAiScreenManager is the usual producer of that scene list - feed it a JSON grid where each line is [时长, 台词, 素材] and it emits the LIST this node expects. In the shipped MiniMax MV workflows, 行索引 comes from a for-loop counter (easy forLoopStart or the pack's FxAIGeneratorController), 台词 feeds the prompt, and 素材 feeds the reference-image input of FxAiMiniMaxImageToVideoV2. Rinse, repeat, stitch.

The inputs that actually matter

• 场景数据 (LIST) - the scene table. Wire this from any node that emits a list of dicts with 台词 and 素材 keys (FxAiScreenManager is the easy path).
• 行索引 (INT) - the current loop iteration, the one that changes each pass; drive it from your loop counter.
• 通用提示词 / 尾部通用提示词 (STRING, optional) - style lock for every scene. Put your global "cinematic, consistent character" instructions here so each scene carries its own dialogue.

Outputs: 台词 (STRING) and 素材 (IMAGE). Nothing else - no frame counts, no audio. If you need those, you've installed the wrong sibling.

Install

ComfyUI Manager, search fxai-toolkit (the display name is "FxAi(凤希全能节点)"), install, restart. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/fxai666/fxai-toolkit
# restart ComfyUI

On first launch the pack auto-installs soundfile and psutil via pip, and that's the whole dependency story - no model files are downloaded by the pack itself. Caveat: this node exists to feed local MiniMax H3 generation, and that part is not free or global - H3's weights are licence-locked out of the US, EU, UK and South Korea, and the workflows expect multi-GB checkpoints from the Comfy-Org MiniMax-H3 repo. The node runs fine without them; your long-video ambitions may not.

Common issues

• Images come back empty. Nine times out of ten the paths in 素材 don't exist under ComfyUI/fxai/image/. The node just logs [凤希] 图片不存在: ... and skips the file - no error, no red. Check the console.
• You looped past the end and re-rendered scene 1. That's the modulo wrap working as intended - great for infinite loops, but if your loop count and scene count disagree you'll silently repeat scenes. Want a hard stop on a bad index? Use the full FxAiScreenLoad instead.
• The whole UI is Chinese. Display names, inputs, README, the works. The input names above translate to "scene data", "row index", "common prompt", "tail common prompt". If you're stuck, the pack's support is the QQ group and the author's Bilibili channel - don't expect an English issue thread to answer.
• Missing soundfile/psutil errors. If the auto-install failed (offline install, weird Python env), the pack's audio nodes will throw at import. This particular node only needs PIL and torch, which stock ComfyUI already ships.