Generate Video (from Start and End Frames)

DeepGen_I2V2 is the image-to-video node that knows where it's going. Where DeepGen_I2V0 animates from a single start frame and lets the model improvise the ending, this one takes a start image and an end image and produces a clip that transitions between them. It's interpolation with a budget: you're not describing a change, you're showing it.

Why you'd reach for it

Any time a transition needs to land somewhere specific. Morph a character from pose A to pose B - same person, clearly different moment. Turn day into night or summer into winter with actual visual bookends. Cut from a product to its packaging, a building to its interior, a sketch to its finished render. Two-frame i2v is the trickiest of the pack's video modes to get right and the most satisfying when it works, because you've taken the guesswork out of the ending.

The model dropdown is the subset that supports start+end frames: the Kling 2.5 Turbo / Turbo Pro family, Kling 3.0, 3.0 Pro and the Omni pair, plus Veo 3.1 and 3.1 Fast. All cloud, all metered, none runnable locally - the usual API-wrapper deal.

How it works

Both images get converted to base64 PNG attachments and submitted with the prompt; the job queues, and the node polls every 15 seconds until the clip is done - expect minutes, with your workflow blocked during the wait. The result downloads to ComfyUI's temp directory and returns as the pack's ComfyVideoMock (its VIDEO-type stand-in), ready for the standard preview and save nodes.

The inputs that matter

• start_image and end_image (IMAGE) - the bookends. Both optional, but the node only earns its name with both connected.
• prompt - describes the motion between the frames: "the camera orbits from the front to the side view." The images carry the state change; the text steers the choreography.
• model - start/end-capable models only.
• duration (1–15, default 5) - seconds, snapped to model support; your main cost dial along with resolution.
• generate_audio (default on) - flip off for cheaper test runs where supported.
• aspect_ratio / resolution - free-form strings, blank = model's choice.
• nb_results, seed_value, output_prefix, config_json - the standard set.

Outputs

VIDEO (the transition clip), output_prefix_and_model (STRING), total_credits_used (FLOAT).

Install

Same pack, standard steps:

cd ComfyUI/custom_nodes
git clone https://github.com/deepiksdev/ComfyUI-DeepGen-API.git
cd ComfyUI-DeepGen-API
pip install -r requirements.txt

Restart and add your DeepGen API key (deepgen.app) in Settings → "DeepGen API Key" (ComfyUI/user/deepgen/config.json).

Gotchas

The classic failure mode is two frames so different the model invents a bizarre intermediate - keep the start and end visually related, and use the prompt to describe the transition, not the endpoints. Your images go off-box, and video minutes burn credits faster than anything else in this pack, so test with short durations and audio off. And note this node has no deprecated twin - the deprecated video names in the pack are DeepGen_T2V (text) and the multi-image siblings, not this one.