Every clip in our feed ends the way a scene should — and this is the exact graph that makes them.
What it does. Drop any picture (yours, a generation, a frame) and get a 4.5–6.5 s live-action video with sound that continues the scene in the first frame and ends in a visible finale: the same people, the same position, a fixed camera, no cuts. Face, mouth, tongue, breasts, belly, back, ass, pussy, inside, thighs or a solo orgasm — 11 tuned finale texts, pick one or let the director choose.
Why it works.
The director graph: two prompts (ANALYZE → DIRECT) that read the picture — who is where, the act, where the cum naturally lands from that position, how long the clip should be — and write the H3 prompt in the form H3 understands (Image 1 lock, second marks, concrete semen, a sound line). Run them in any vision LLM, or let the optional Ollama group run them inside ComfyUI.
The LoRA stack: Cum 0.75 + Unlocked 0.5 lead, the position LoRA of the act (Deepthroat/Face Fuck for oral, riding, VBVR, Prone Bone…) as the third slot. Tuned on hundreds of clips: the semen is actually there in the last second.
One position per clip — the rule that killed our broken clips. Baked into the rules tail.
Universal. Any H3 checkpoint, any picture style (photo, 3D, cartoon — the clip keeps the style of the frame), any aspect (portrait/landscape/square), 8 GB VRAM at 0.5 MP with the 4-step turbo.
Simple. Core ComfyUI nodes only — nothing to install for the main path. Four groups, titled nodes, every setting a visible widget, README and model list inside the graph as notes. Open, read, edit.
Includes: the workflow (UI + API JSON), the director prompts and finale texts as text, README with the model list and the quality rules.
Made with the same pipeline that runs our account every day — if you like the feed, this is how it is made.