LTX 2.5 / firstLastFrameToVideo

Image-to-video usually means "start here, then improvise." This node is the version that won't improvise the landing: you give it a first_frame and/or a last_frame, and the model has to hit both. That's how you make loops, transitions, or a shot that ends exactly where the next scene begins - the bookends are fixed, everything between is the model's problem. The required input list is deceptive here: the schema only requires prompt, but the tooltip on last_frame is explicit - at least one of FirstFrame or LastFrame must be provided or the job isn't going to do what you think. The frames are STRING fields taking a URL, data URL, or base64, not IMAGE sockets.

Part of civitai-comfy-nodes, Civitai's official ~160-node pack generated from its Orchestration API spec. Cloud job: LTX 2.5 runs on Civitai's fleet through Comfy workflows, billed in Buzz. Submit, long-poll, get a native VIDEO tensor.

Inputs that matter

• prompt (required) - what happens between the bookends.
• first_frame / last_frame (STRING) - one or both required in practice. A URL, data URL, or base64 image. Provide both for a hard loop or transition; just one to anchor only that end.
• frame_guide_strength - default 0.7, 0.0–1.0. How hard the model clings to the guide frames. Lower it if the output's trying too hard to freeze at the bookends.
• duration (3–20s, default 5), width/height (1280×720), fps (24), seed, guidance_scale (4), steps (20), generate_audio (on) - the usual LTX 2.5 surface.
• model (22b-dev), loras, diffusion_model (fine-tune override), quantity.

Outputs: video (VIDEO), draft_cache_audio (AUDIO), draft_cache_video (VIDEO), workflow_id and raw_json (STRING).

Install and auth

Standard for the pack - official, Comfy Registry:

# ComfyUI Manager: search "Civitai Comfy Nodes"
cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt

Restart, authenticate via a Civitai Auth node, CIVITAI_API_TOKEN, or sidebar OAuth.

Common issues

The STRING-vs-socket trap is the big one. first_frame and last_frame are text fields for URLs/base64, not IMAGE sockets - people wire a Load Image in expecting it to connect, and it won't. Either paste a URL or convert your image to a data URL first. Forgetting both frames leaves you with a prompt-only generation despite the node's name; one frame at minimum. Loop stitching is finicky - perfect loops need the two frames to be visually consistent (same camera, same lighting) or the seam shows. frame_guide_strength at 1.0 can stall the motion near the bookends; 0.7 is the sane default. Your frames leave your machine, and the pack is early preview - behavior can change without notice.