All workflows
Every workflow is a ComfyUI file that loads as a finished graph and uses only ComfyUI’s built-in nodes. Open it with File → Open, or drag it onto the canvas.
Download this guide as Markdown (for your AI agent)
Version 1 (2026-10-15): the same steps as this page, with every download’s link and checksum. sha256 555f1460ccb4d9d81e8a03eeedee875cf49fbb70ad1aa72ae6873214b18a70a6, also in the manifest.
You need ComfyUI v0.39.0 (version note) and the models on Set up ComfyUI.
“16 GB” means a 16 GB GPU with 32 GB of system RAM and a pagefile of about 32 GB or more. “48 GB” means a 48 GB-class GPU or a rented cloud GPU. Every figure was measured on a 16 GB RTX 5070 Ti with 32 GB of RAM unless it says otherwise.
Swap a person
h3-ref2va-character-swap-blank-reference.16gb.json
- Does
- Recasts one person in the whole shot: they’re greyed out in the reference video only; the reference is padded to the model’s grid and the original sound is kept (and a
-4x3version). Also changes what someone wears, with the garment greyed instead. Validation pending: its GPU check on our side is still to come. - Guide
- Swap a person
- GPU
- 16 GB; about 3¼ minutes for a 124-frame part at 704 × 544; longer parts take roughly 3 seconds per frame. Parts of 175 frames or more use almost all of 32 GB RAM. Close other apps (game launchers, browsers) before you run it, or ComfyUI can be killed partway through.
h3-ref2va-character-swap-blank-reference.16gb-4x3.json
- Does
- The same, for 4:3 shows (768 × 576 as shipped).
- Guide
- Swap a person
- GPU
- 16 GB; about 3¼ minutes for a 124-frame part at 704 × 544. Uses almost all of 32 GB RAM. Close other apps (game launchers, browsers) before you run it, or ComfyUI can be killed partway through.
seamless-people-masks-1.json
- Does
- Marks one person you name on every frame of your shot, for the grey reference: the person you recast, or the wearer for an outfit change. Saves them, them grown by 8, and the background around them. Never leave its word blank: a blank finds the whole person. Built-in nodes only.
- GPU
- 16 GB; about a minute for a 6-second shot.
seamless-people-masks-2.json
- Does
- Marks the people or things you name on every frame of your shot, in one run: saves them, them grown by 8, and the background around them. Download the version for your number of words:
-1,-2,-3or-6.json(fill every slot: repeat your first word in any you don’t need, because a blank slot marks the whole person). Built-in nodes only. - GPU
- 16 GB; about a minute per person for a 6-second shot (3 people: 155–201 s).
seamless-people-masks-3.json
- Does
- The same, for three people or things. Repeat your first word in any slot you don’t need: a blank slot marks the whole person.
- GPU
- 16 GB; about 3 minutes for a 6-second shot (155–201 s).
seamless-people-masks-6.json
- Does
- The same, for up to six people. Repeat your first word in any slot you don’t need: a blank slot marks the whole person.
- GPU
- 16 GB; about a minute per slot for a 6-second shot.
seamless-track-one-person.json
- Does
- Follows one person you box on the first frame through the whole shot, for the grey reference when words pick the wrong people (an entry, someone passing in front). Saves the person, them grown by 8, and a red preview over your shot. For a person who walks in, reverse the shot first (the guide shows how). Built-in nodes only. Validation pending: its GPU check on our side is still to come.
- GPU
- 16 GB; a few minutes for a 6-second shot (not yet measured on our side).
seamless-mask-keep-on-person.json
- Does
- Keeps only the part of a mask video that falls on one person (their mask, grown a little): for example, a garment-and-arms mask kept to the wearer. White = greyed out in the reference only; nothing is pasted back.
- GPU
- 16 GB; about 10 seconds (5–20).
seamless-mask-keep-inside-box.json
- Does
- Keeps only the part of a mask video that falls inside a box you place (left, top, width, height, and the frames it covers): for example, a garment mask kept to the wearer when the wearer’s own mask drops out or jumps (steady shots). White = greyed out in the reference only; nothing is pasted back.
- GPU
- 16 GB; about 10 seconds.
seamless-box-mask.json
- Does
- Makes a mask video with a box you place (left, top, width, height, and the frames it covers), for the grey reference when a word mask jumps to the wrong person. White = greyed out in the reference only; nothing is pasted back.
- Guide
- Swap a person, Check
- GPU
- 16 GB; a few seconds.
h3-ref2va-character-swap.16gb.json
- Does
- The swap without a grey reference, for a non-human character or a short part. 1024 × 576, 124 frames, 4 steps.
- Guide
- Swap a person
- GPU
- 16 GB. Peak 13.2–15.1 GiB GPU, 31.2 GB RAM; 4–7 minutes per clip.
h3-ref2va-character-swap.16gb-4x3.json
- Does
- The same swap for 4:3 shows, at 768 × 576.
- Guide
- Swap a person
- GPU
- 16 GB; about 3 minutes per clip.
h3-ref2va-character-swap.hq-48gb.json
- Does
- The higher-quality swap, 1536 × 864, 192 frames.
- Guide
- Swap a person
- GPU
- 48 GB. It ran out of memory on 16 GB.
h3-ref2va-character-swap-extra-ref.16gb.json
- Does
- The plain swap workflow with an extra picture input for a person who stays (and a
-4x3version): load a crop of them in Extra reference: a crop of a person who stays. - GPU
- 16 GB; about the same as the plain swap: about 3¼ minutes per 124-frame part at 704 × 544.
h3-ref2va-character-swap-extra-ref.16gb-4x3.json
- Does
- The same, for 4:3 shows (768 × 576).
- GPU
- 16 GB; about the same as the plain swap: about 3¼ minutes per 124-frame part at 704 × 544.
Add new footage
h3-fl2va-first-last-frame.16gb.json
- Does
- A new clip that starts on one frame and ends on another. The seamless extension uses it to leave the episode’s picture and come back to it. 1024 × 576, 124 frames, 8 steps.
- Guide
- Add new footage
- GPU
- 16 GB; about 2½ minutes per clip (about 1½ minutes at 4 steps for drafts).
h3-fl2va-first-last-frame.hq-48gb.json
- Does
- The higher-quality version of the above.
- Guide
- Add new footage
- GPU
- 48 GB.
h3-i2v-first-frame.16gb.json
- Does
- A new clip from a first frame only: a cutaway, flashback or vision, or a moment that leaves the episode and goes somewhere new. 1024 × 576, 124 frames.
- Guide
- Add new footage
- GPU
- 16 GB; about 3 minutes per clip.
h3-i2v-first-frame.hq-48gb.json
- Does
- The higher-quality version of the above.
- Guide
- Add new footage
- GPU
- 48 GB.
h3-fl2va-last-frame-only.16gb.json
- Does
- Makes new footage that ends exactly on a frame of your episode, with the start left free: for a moment that comes from somewhere and lands on the episode (and a
-4x3version). - GPU
- 16 GB; about 3¼ minutes for 124 frames at 1024 × 768 (one run, including loading the model).
h3-fl2va-last-frame-only.16gb-4x3.json
- Does
- The same, for 4:3 shows (1024 × 768).
- GPU
- 16 GB; about 3¼ minutes for 124 frames.
h3-ref2v-pictures-start-frame.16gb.json
- Does
- Makes new footage that starts on a frame of your episode and brings in what your pictures show (a person, a place, an object): for added context (and a
-4x3version). - GPU
- 16 GB; about 2¼ minutes for 124 frames at 1024 × 768.
h3-ref2v-pictures-start-frame.16gb-4x3.json
- Does
- The same, for 4:3 shows (768 × 576 as shipped).
- GPU
- 16 GB; about 2¼ minutes for 124 frames.
h3-ref2v-pictures.16gb.json
- Does
- Makes new footage from your pictures and the prompt with neither end fixed: for a cutaway that fades in and out (and a
-4x3version). - GPU
- 16 GB; about 2 minutes for 124 frames at 1024 × 768.
h3-ref2v-pictures.16gb-4x3.json
- Does
- The same, for 4:3 shows (768 × 576 as shipped).
- GPU
- 16 GB; about 2 minutes for 124 frames.
seamless-add-scene-finish.json
- Does
- Finishes new footage for a 16:9 show: covers and center-crops it to the Start still size, sets the sound level, and resamples the sound to 48 kHz at exactly the picture’s length. For a 4:3 show, use
seamless-add-scene-finish-4x3.json. Built-in nodes only. - GPU
- About 10 seconds.
seamless-add-scene-finish-4x3.json
- Does
- The same finish for a 4:3 show, at 1024 × 768. Pick the file by the episode’s shape: the 16:9 file cover-crops a 4:3 shot.
- GPU
- About 10 seconds.
Rebuild a shot
seamless-upgrade-effects-fullframe.16gb.json
- Does
- Redraws the whole frame with a new design, guided by the depth of your original shot, so the new version moves exactly like the old one; its look comes from your reference pictures. For a new creature, ship or effect. MiniMax H3 Ref2VA with the Fun ControlNet depth patch, Turbo v4 LoRA, 8 steps, control strength 1.5, 1024 × 576.
- Guide
- Rebuild a shot
- GPU
- 16 GB. Measured: a 345-frame shot at 768 × 576 in one run, about 13 minutes.
seamless-upgrade-effects-fullframe.16gb-4x3.json
- Does
- The same, at 1024 × 768 for 4:3 shows.
- Guide
- Rebuild a shot
- GPU
- 16 GB. 1024 × 768 is the card’s ceiling: system RAM peaked at up to 32.4 of 32 GB. Close every other app.
seamless-upgrade-effects-fullframe-actors.16gb.json
- Does
- The whole-frame redraw with your shot also going in as a small reference video, so the shot’s motion is held exactly: for extras, crowds and views where the design stays. The reference video carries the old design, so for a new design use the file above. 1024 × 576.
- Guide
- Rebuild a shot
- GPU
- 16 GB, tested. 5 seconds at 1024 × 576: about 4½ minutes, 14.6 GB GPU, 28.6 GB RAM.
seamless-upgrade-effects-fullframe-actors.16gb-4x3.json
- Does
- The same, at 1024 × 768 for 4:3 shows.
- Guide
- Rebuild a shot
- GPU
- 16 GB, tested. 5 seconds at 1024 × 768: about 6½ minutes, 14.4 GB GPU, 28.9 GB RAM.
seamless-cutout.json
- Does
- Cuts the thing you name out of a reference picture: one PNG on plain grey, one with transparency.
- GPU
- About 5 seconds per picture.
Rebuild a shot with actors
seamless-upgrade-effects-fullframe-actors-blank-reference.16gb.json
- Does
- Rebuilds a shot in full while the actors stay themselves: what you replace (a creature, the place) is greyed out in the reference video only, and the actors stay visible there (and a
-4x3version). The output is generated in full; nothing is pasted back. - GPU
- 16 GB; about 6 minutes for 124 frames at 1024 × 768. Split long shots into parts of about 124 frames; very long parts can run out of memory on a 32 GB machine. Uses almost all of 32 GB RAM. Close other apps (game launchers, browsers) before you run it, or ComfyUI can be killed partway through.
seamless-upgrade-effects-fullframe-actors-blank-reference.16gb-4x3.json
- Does
- The same, at 1024 × 768 for 4:3 shows.
- GPU
- 16 GB; about 6 minutes for 124 frames. Uses almost all of 32 GB RAM. Close other apps (game launchers, browsers) before you run it, or ComfyUI can be killed partway through.
seamless-upgrade-effects-fullframe-actors-blank-reference-6pics.16gb.json
- Does
- The rebuild with actors workflow with three more picture inputs, for a face crop of each actor who stays (and a
-4x3version). For shots with three or more actors, or a new creature with actors. The output is generated in full; nothing is pasted back. - GPU
- 16 GB, tested. 124 frames at 1024 × 768: about 6 minutes (5½–6½ over 19 runs). Uses almost all of 32 GB RAM. Close other apps (game launchers, browsers) before you run it, or ComfyUI can be killed partway through.
seamless-upgrade-effects-fullframe-actors-blank-reference-6pics.16gb-4x3.json
- Does
- The same, at 1024 × 768 for 4:3 shows.
- GPU
- 16 GB, tested; about 6 minutes for 124 frames. Uses almost all of 32 GB RAM. Close other apps (game launchers, browsers) before you run it, or ComfyUI can be killed partway through.
seamless-effects-mask.16gb.json
- Does
- Marks the thing you replace on every frame of your shot and saves a black-and-white mask video, for the grey reference. Type a word for it (
dinosaur,spaceship,monitor); never leave it blank, because a blank finds the whole person. Every look-alike in a frame is included, and the mask is grown for room (12). If one word misses a part, use the two- or four-word version. - GPU
- 16 GB; about a minute for a 5-second shot.
seamless-effects-mask-2words.json
- Does
- The mask workflow with two words (and
seamless-effects-mask-4words.jsonwith four), for when one word misses part of the thing you replace. Every word is marked and the masks are joined. Fill every slot: repeat the first word if you don’t need a second (a blank slot adds the whole nearest person to the mask). - GPU
- 16 GB; about 1¾ minutes for a 5-second shot (2 words), 2–2½ (4 words).
seamless-effects-mask-4words.json
- Does
- The mask workflow with four words. Every word is marked and the masks are joined. Fill every slot: repeat your first word in any slot you don’t need (a blank slot adds the whole nearest person to the mask).
- GPU
- 16 GB; about 2–2½ minutes for a 5-second shot.
Workflows every guide shares
seamless-silent-track.json
- Does
- Adds a silent 48 kHz sound track to a clip that has none, so the workflows that use the shot’s sound as an anchor can run.
- Guide
- Capture, sound
- GPU
- About 5 seconds.
seamless-crop-4x3.json
- Does
- Center-crops a 16:9 recording of a 4:3 show to 4:3 and scales it to 1024 × 768, sound kept.
- Guide
- 4:3 shows
- GPU
- About 10 seconds.
seamless-still-1024x576.json
- Does
- Scales a still to cover 1024 × 576 and crops the overflow from the center. Also for 1024 × 768 and 768 × 576 (
seamless-still-1024x768.json,seamless-still-768x576.json). - Guide
- Add new footage; 4:3 shows
- GPU
- About 10 seconds.
seamless-still-1024x768.json
- Does
- The same, at 1024 × 768 (a 4:3 show).
- Guide
- 4:3 shows
- GPU
- About 10 seconds.
seamless-still-768x576.json
- Does
- The same, at 768 × 576 (the 4:3 swap size).
- Guide
- 4:3 shows
- GPU
- About 10 seconds.
seamless-frame-export.json
- Does
- Saves a clip’s first and last frames as PNGs: the next part’s picture, or a move’s end frames.
- GPU
- About 5 seconds.
seamless-finish.json
- Does
- Finishes a clip: keeps your shot’s frame count, plays every frame at your shot’s rate, scales to the Start still and puts your shot’s sound back. Built-in nodes only; no keyframe control.
- Guide
- Finish and conform
- GPU
- About 20 seconds (21 s for a 6-second shot).
seamless-flicker-check.json
- Does
- A check, not an edit: paints red wherever your result flickers against your original, frame to frame. Red blobs on scenery or people are a problem; a thin red line along the moving new element is normal. Built-in nodes only.
- Guide
- Check your clip
- GPU
- About 2 minutes for a 13-second shot (7 GB GPU). Run it on one shot at a time; for shots over about 250 frames close other apps: it needs most of 32 GB of RAM.