凤希AI - LTXV多帧引导器

FxAiLtxvGuideFrames (凤希AI - LTXV多帧引导器) is the keyframe injector for LTX video that ComfyUI's stock LTXVAddGuide node only kind of is. Stock LTX conditioning gives you first-frame and last-frame control; anything in between you're normally left to hope the model drifts there on its own. This node lets you feed a batch of guide images and stamp each one into the conditioning at a specific frame index, mid-clip, with per-guide strength. For long LTX shots where you want a character to look one way at frame 10 and the camera to have moved by frame 80, that's the difference between praying and planning.

Mechanically it's a batched wrapper over ComfyUI's own LTX keyframe machinery - it imports LTXVAddGuide and _append_guide_attention_entry from comfy_extras.nodes_lt, so it's leaning on the same primitives the native "add guide" node uses, just looped. For each image in 引导图批量 (IMAGE batch) it: encodes the image through the 视频VAE at the latent's dimensions, resolves where in the latent timeline the frame should land (via 指定帧索引, in frame space), asserts it's inside the clip, then appends it as a keyframe - writing the guide's latent into the video latent at that position and adding an attention entry so the cross-attention actually references it. The noise_mask is updated alongside, so the guided region is respected during sampling.

The inputs that matter:

• 正向条件 / 负向条件 (CONDITIONING) and 视频潜变量 (LATENT) - the normal LTX encode outputs you're modifying.
• 视频VAE - the LTX video VAE used to encode guide images.
• 引导图批量 - one or more images to plant. Batch of 3 with frame indices 0, 40, 80 and you've got a storyboard.
• 指定帧索引 - which frame each guide lands on (applies per image as it processes the batch).
• 引导强度 - per-guide strength, default 1.0.

Outputs: 正向条件, 负向条件, and the updated 视频潜变量, ready to feed a sampler. It returns the conditioning and latent so the whole "encode → guide → sample" chain can be re-wired in one node.

This is an LTXV-era pack category, and the LTX story still holds: speed-first drafting, weaker on faces/complex motion, and it rewards longer prompts. A mid-clip keyframe is the standard trick for keeping a scene on rails when LTX would otherwise wander.

Install: ComfyUI Manager → search "fxai-toolkit", or cd ComfyUI/custom_nodes && git clone https://github.com/fxai666/fxai-toolkit, restart. Auto-installs soundfile/psutil, FFmpeg on PATH, and everything is labeled in Chinese - 凤希AI builds for a Chinese-speaking community (QQ group 775649071, Bilibili space.bilibili.com/1242183138).

The failure mode to know: it asserts the guide frame fits inside the latent (引导帧超出视频长度范围 if not), so a 指定帧索引 past your clip's frame count is a hard error, not a silent crop. And guide images should roughly match the latent's resolution - the node scales what you give it, but extreme aspect mismatch makes the planted frame look mushy. Keep your batch aligned to your storyboard and this node makes long LTX clips feel scripted instead of improvised.