Load the original graph
Save the workflow PNG to your computer and drag it into ComfyUI, or open the JSON. The diagrams on this page are explanatory HTML mockups. Download the actual file above to run the graph.
MINIMAX H3 + COMFYUI / CONTINUOUS SHOT WORKFLOW
FREE + Local
Made on my PC.
Three character sheets. One tiny photo booth. A cat with absolutely no sense of timing. Here’s how I turn those ingredients into a hand-drawn anime moment with my custom Krea 2 Turbo fine-tune, MiniMax H3 and ComfyUI.
01 / THE FIELD GUIDE
Start with the references and graphs used for this scene. Make one successful test before you build your own world.
The downloads contain graph settings and references, not model weights. My image stage uses a custom trained / fine-tuned Krea 2 Turbo model. You need that checkpoint to reproduce its look; the approved reference PNGs let you start directly at the animation stage. No public download for my custom weights is included here.
Save the workflow PNG to your computer and drag it into ComfyUI, or open the JSON. The diagrams on this page are explanatory HTML mockups. Download the actual file above to run the graph.
Open Manager and use Install Missing Custom Nodes (or the missing-node filter in your version). Install the required packs, then restart ComfyUI and refresh the page. This video graph uses ComfyUI-KJNodes and Comfyui-MMH3-UltimateUpscale.
Use Manager’s Model Manager / Install Models where available to find and download missing weights. Match the exact loader filenames below. If a custom or quantized model is not listed, obtain it from its author and place it in the matching model folder. Manager cannot supply unpublished fine-tunes.
Choose your model files in each loader and select your own files in all four Load Image nodes. Update ComfyUI if core MiniMax nodes are missing. Run a short test with a shortened prompt first; expand to the full scene after the graph works.
Manager absent? Follow the official installation guide ↗. Labels differ between Manager versions; its project documentation ↗ covers node and model management.
| Stage / folder under ComfyUI/models | Selected file |
|---|---|
| Images / diffusion_models | Wulver_v0.5_fp8_e4m3fn.safetensors |
| Images / text_encoders | qwen3vl_4b_bf16.safetensors |
| Images / vae | qwen_image_vae.safetensors |
| Video / diffusion_models | minimax_h3_fused_refdelta_r1024_turbo8_mystic07_int8_convrot.safetensors |
| Video / text_encoders | qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors |
| Video / vae | minimax_h3_video_vae_int8_convrot.safetensors |
| Audio / vae | minimax_h3_audio_vae_fp32.safetensors |
| Upscale / latent_upscale_models | minimax_h3_latent_upscaler_3d_fp16.safetensors |
The character PNGs record the image checkpoint as wulverKrea2_v05_fp8.safetensors; the supplied starter JSON uses the filename above. Select your installed compatible Wulver checkpoint. A filename alone does not establish that two files have identical weights.
The video VAE loader selects the INT8 file, while its saved download hint points at an FP16 VAE. Follow the selected loader value and compatibility requirements, not an old hint. Quantized variants may require their supporting ComfyUI build. These files were inspected; a fresh model installation and render have not been tested for this tutorial.
02 / THE FIELD GUIDE
Give the video model a clear visual identity before you ask it to act.
I use my custom Krea 2 Turbo fine-tune to establish the look: chunky organic ink lines, painterly cel shading, warm paper tones, teal clothing and orange accents. Generate each character separately. The three views describe one individual, not three people to place in the scene.

Character sheet of a cheerful 30 year old young Asian man, wearing thick black framed glasses, slim build, messy short black hair, large expressive dark eyes, rosy cheeks. Pale blue T-shirt, black pants, blue Converse sneakers with white toe caps and laces. No apron or overalls. Three views arranged horizontally: LEFT three-quarter close-up portrait, MIDDLE full-body front, RIGHT full-body back. Consistent proportions and outfit, feet fully visible, plain cream background. Hand-drawn 2D anime, chunky organic outlines, expressive cartoon face, painterly cel shading, textured handmade brushwork, vibrant teal-turquoise-orange accents, animated-film concept art. No text, photorealism, or 3D.

Character sheet of a cheerful young 28 year old Asian woman, long messy dark-brown hair with bangs, large expressive dark eyes, rosy cheeks. Red T-shirt, teal blue-green denim overalls with rolled cuffs, faded red Converse sneakers with white toe caps and laces. No apron. Three views arranged horizontally: LEFT three-quarter close-up portrait, MIDDLE full-body front, RIGHT full-body back. Consistent proportions and outfit, feet fully visible, plain cream background. Hand-drawn 2D anime, chunky organic outlines, expressive cartoon face, painterly cel shading, textured handmade brushwork, vibrant teal-turquoise-orange accents, animated-film concept art. No text, photorealism, or 3D.

Wide character sheet of a cute chubby orange tabby cat: round head, plump compact feline body, very short legs, small black oval eyes, pink nose, cream muzzle and chest, darker orange stripes, thick curved tail, teal collar with a small teal locket. Three clearly separated views with generous space: LEFT three-quarter head-and-shoulders close-up; MIDDLE full-body front, sitting naturally with front paws together; RIGHT full-body back in the same seated pose, tail curled beside the body. Small ears, rounded cheeks, consistent proportions. Plain cream background. Hand-drawn 2D anime, chunky organic outlines, painterly cel shading, textured handmade brushwork, warm orange and teal palette. No human anatomy, upright human stance, long limbs, overlapping figures, clothing, text, or 3D.
The close-up establishes face and hair. The front view establishes proportions, clothes and footwear. The rear view gives the model information for turning and entering the booth. A cream background separates the silhouette from scenery.
Keep the same palette and rendering vocabulary across all three prompts. Precise clothes and markings give you concrete things to inspect when choosing a result.
The starter graph is text-to-image: model + text encoder + VAE → text conditioning → sampler → decode → save. It has no image-reference conditioning path. Put the descriptions in the positive prompt.
The supplied starter uses 1024 × 1024, 8 steps, CFG 1.0, Euler, simple, denoise 1.0. Kai and Niko’s approved PNGs record 1280 × 1280. Set the latent dimensions deliberately. Keep the negative field empty at this CFG.
Custom fine-tune
Text encoder + VAE
Identity → outfit → views → style
8 steps · CFG 1
Euler / simple
Approved character PNG
The expandable prompts above are the actual positive prompts embedded in the supplied character PNGs. Check glasses, straps, shoes, tabby markings and body proportions before approving a sheet.
03 / THE FIELD GUIDE
A location reference should solve composition before the performance begins.

“A cute photo booth” describes a category. “A narrow worn bench, viewed from the booth camera at seated eye level” establishes where people sit and where the viewer is. Side windows leave room for background life without competing with the main action.
Empty means the bench is empty. Pedestrians can remain outside. Later, the prompt keeps walls and furniture stationary while people continue walking through the windows.
The overall image workflow uses my custom Krea 2 Turbo fine-tune. This particular booth PNG also records a Qwen Image 2.1 cleanup/edit pass: remove the characters, cat and photostrip, then add a worn faux-leather bench. That separates the setting from the performance.
The prompt below is a new text-to-image reconstruction for the Krea workflow, not the exact edit prompt that produced this PNG.
One continuous square illustration of an empty retro station photo booth, viewed from its built-in camera at seated eye level. A cramped, weathered teal interior, worn brown faux-leather bench, warm striped overhead light, small wall signs, side windows opening onto a station concourse. Keep the bench empty and leave space for two seated adults. Small secondary pedestrians are visible outside the booth through the side windows. Hand-drawn 2D anime, chunky organic outlines, painterly cel shading, textured handmade brushwork, warm teal and orange palette. No people or animals inside the booth, no photostrip, collage, captions, photorealism or 3D.
04 / THE FIELD GUIDE
The scene is a tiny comedy with a setup, a pattern, an interruption and a payoff.
SCENE 01 / INT. STATION PHOTO BOOTH
An empty booth waits. Kai and Niko squeeze onto the bench. Finger hearts. Flash. Peace signs. Flash. Mochi springs into frame. Their smiles turn to surprise. Flash. They pull the cat close and grin into the lens. One last flash.
ONE CAMERA · THREE CHARACTERS · FOUR READABLE MOMENTSThe first two photos establish a predictable rhythm: pose, hold, flash. Mochi breaks that rhythm. The final group photo resolves the surprise with affection. You can understand the scene without dialogue because the expressions carry the change.
Each action ends in a pose that can be photographed. “They make a peace sign” can keep moving indefinitely. “Finish the gesture, hold it, then flash” gives the model a sequence and a stable target. Timestamps are direction, not frame-accurate promises.

Finger hearts. Hold the gesture before the flash at 5.2s.

Peace signs. A second hold and flash at 7.7s. Mochi stays hidden.

A sudden pop from below. Keep the surprise through the 10.4s flash.

Reach for Mochi. Bring everyone together. Final flash at 14.2s.
Stills are actual frames selected from the finished clip. The times describe the intended prompt schedule; generated timing and gestures can vary. The opening 0–3 seconds establish the booth and bring the couple onto the bench.
05 / THE FIELD GUIDE
Use reference-to-video to compose a performance from identities and a location.

IMAGE 1 · Kai

IMAGE 2 · Niko

IMAGE 3 · Mochi

IMAGE 4 · Booth
Identity + style + staging
Timed action → pose hold → flash
Reference-to-video
Euler + manual sigmas
MMH3 Ultimate Upscale
1024 × 1024 · 1 refinement step

24 FPS EXPORT / SQUARE
Kai feeds reference image 0, Niko image 1, Mochi image 2 and the booth image 3 on node #108. Those are Images 1–4 in the prompt. Keep this order, or change both the wiring and the descriptions together.
The finished group image is an output example. It is not a fifth required reference and not a first-frame input. Reference-to-video creates a composition; it does not guarantee the booth image will be copied as the opening frame.
The text comes from Reference Prompt #103. A connected prompt overrides old text still stored inside the reference node. Likewise, the resolution and frame count come from the connected selector and expression nodes.
The supplied graph contains old node titles and leftover internal text. Do not copy those into the tutorial prompt. Follow the actual wires and selected values described here.
06 / THE FIELD GUIDE
Tell the model who is present, what moves, what stays fixed and when the action resolves.
“Image 1 defines Kai’s identity and clothing.” Repeat for Niko, Mochi and the booth. This separates identity from setting. Add “the sheets show multiple views of the same individual” to discourage the model from treating every panel as another character.
“1990s anime” is a broad style label. Stepped drawings, sparse in-betweens, expressive key poses and deliberate holds describe what motion should look like. Small anticipation and hair follow-through keep limited animation expressive.
A fixed camera does not mean everything is frozen. Lock the booth architecture and framing, then explicitly allow station pedestrians to walk. Specify Kai on screen-left and Niko on screen-right to keep staging readable.
Mochi pops up; the couple flinches and drops their signs; the surprised pose settles; the flash captures it. This causal order is more usable than listing “cat, surprise, flash, cute” as disconnected keywords.
Keep surprise through the third flash. Begin smiling only afterward. Reserve a separate final beat for the cuddle and fourth photograph. Otherwise the emotional payoff can start too early and erase the joke.
One continuous shot, no reframing, no duplicate cats, brief flashes, no dialogue. These constraints address likely failure modes. They work best after a clear positive description of the performance.
The supplied workflow’s prompt still contains the slow peek in the surprise beat and the revised sudden entrance under the final-photo heading. This version consolidates your later direction: one sudden photobomb at 8.0–10.7s, then the group cuddle at 10.7–15.0s. The original downloadable PNG and JSON remain unchanged. Paste this corrected prompt into node #103 before running.
15-second video, square 1:1 composition. One continuous static shot from the photo booth’s built-in camera, facing the bench at seated eye level. REFERENCE ROLES Image 1 defines Kai’s identity and clothing: messy short black hair, large black glasses, light blue T-shirt, black trousers. Image 2 defines Niko’s identity and clothing: long dark brown hair with bangs, red T-shirt beneath teal denim overalls. Image 3 defines Mochi: a small, round orange tabby cat with darker orange stripes, cream muzzle and chest, teal collar and round teal tag. Image 4 defines the booth’s exact layout, framing, perspective and painted background: cramped teal interior, worn brown bench, overhead light, side windows overlooking the station. Preserve this setting throughout. The character sheets show multiple views of the same individual. Render exactly one Kai, one Niko and one Mochi. ANIMATION STYLE 1990s hand-drawn 2D cel animation with the visibly stepped cadence of approximately 15 distinct drawings per second. Strong expressive key poses, sparse in-between drawings, deliberate pose holds, small anticipation movements and brief hair follow-through. Hand-inked outlines, flat cel shading, warm painted textures and a softly painted background matching the references. Hold drawings still during each photo pose. Keep the camera and booth architecture stable; allow background pedestrians to move. CAMERA AND BACKGROUND Keep the camera completely fixed throughout. The booth’s walls, bench, windows, signs and station architecture remain stationary, preserving Image 4’s composition. Through the side windows, station pedestrians walk naturally past the booth with the same stepped, hand-drawn animation cadence as the main characters. Their movement continues throughout the shot, including during the photo pose holds. Keep them secondary to the foreground action. STAGING Kai sits on screen-left and Niko on screen-right, squeezed shoulder-to-shoulder on the narrow bench. Their faces and posing hands remain clearly visible. They look directly into the booth lens for each photograph. Mochi enters from below the bottom-center edge only after the second photograph. TIMELINE [0.0–1.0 seconds] The booth bench is empty. Hold on the unoccupied booth. [1.0–3.0 seconds] Kai enters from screen-left and Niko from screen-right. They quickly shuffle onto the narrow bench, bump shoulders gently and scoot close together. Their movement settles with a small bounce and a little delayed hair movement. They exchange an affectionate glance, then face the lens. [3.0–5.5 seconds — first photograph] They lean their heads toward each other and each raises one hand beside their own cheek, making a small thumb-and-index-finger heart. Both smile warmly at the lens. Finish forming the gestures by 4.0 seconds, then clearly hold the completed pose. At 5.2 seconds, one brief bright photographic flash fires from the camera position. They remain in the same pose through the flash. [5.5–8.0 seconds — second photograph] They lower the finger hearts and change to playful V-shaped peace signs beside their faces. Niko tilts her head toward Kai; Kai gives a broad, cheerful grin. Finish the pose by 6.5 seconds and hold it still. At 7.7 seconds, a second brief photographic flash fires. Mochi is still entirely out of sight. [8.0–10.7 seconds — sudden photobomb photograph] Immediately after the second photograph, Mochi suddenly pops up from below the bottom-center edge in one quick, springy motion, photobombing the couple. The cat’s ears, face and upper body appear together, filling the lower-center foreground between them. No slow peek or gradual reveal. Kai and Niko instantly flinch backward in surprise, dropping their peace signs, widening their eyes and opening their mouths in startled expressions. Mochi faces directly into the lens with an innocent, cheerful expression. Hold this surprised group pose clearly before the third flash fires at 10.4 seconds, capturing the photobomb. Keep their startled expressions through the flash; they begin smiling and reaching for Mochi only afterward. [10.7–15.0 seconds — final photograph] Their surprise softens into delighted smiles. Together they gently scoop Mochi upward and cradle the cat between them at chest height, supporting its body. They lean inward and hug Mochi together, cheeks close to the cat’s head. Mochi remains a small, round cat, comfortably nestled in their arms, with a contented smiling expression. All three face the lens. Complete the final group pose by 13.0 seconds and hold it. At 14.2 seconds, the fourth and final flash fires. Normal exposure immediately returns, revealing the same happy group pose, held until the video ends. FLASH BEHAVIOR Exactly four brief flashes, one for each completed photo pose. Each flash momentarily washes the existing scene toward white and immediately returns to normal exposure. Every flash preserves the shot’s composition, character positions and continuity. The pose is established and held before the flash fires. AUDIO Soft booth room tone, gentle clothing rustle, a small bench creak and four shutter clicks synchronized with the flashes. No dialogue, talking, lip-sync, vocalizations or music. CONSTRAINTS No cuts, zooms, pans, camera shake or reframing. No continuous strobing or prolonged white screen. No additional foreground characters or duplicate cats. Preserve Kai’s glasses and every character’s reference clothing and markings. No character-sheet panels, photo-strip overlays, captions or added graphics. No 3D rendering, photorealism, fluid high-frame-rate motion, frame blending or motion-interpolation appearance.
Download the corrected video prompt ↓
This structure builds on the project’s original prompting discussion and fal’s MiniMax H3 prompting guide ↗. The explanations here are specific to this scene, not a guarantee of exact generated timing.
07 / THE FIELD GUIDE
Use the saved graph as a starting point and judge the action before spending time on a final pass.
The active base path uses Euler and ManualSigmas with 0.9999166, 0.9728326, 0.9230769, 0.8, 0.0: four sampling intervals. The old BasicScheduler is bypassed. Changing its “steps” field will not control the active base sampler.
The upscale path uses the dedicated H3 latent upscaler, sa_solver, and a simple scheduler with 1 step / denoise 0.2. Keep the downloaded graph’s connections intact while learning the flow.
The duration expression rounds the requested seconds onto a 17k + 5 frame grid. At a 15-second target and 24 fps, it requests 362 frames, approximately 15.08 seconds.
“About 15 distinct drawings per second” describes the visual cadence. The file is assembled at 24 fps. Setting export fps to 15 alone would slow the same frames down and change the timing; it would not create better limited animation.
Node #7 is set to 1:1, 0.4 MP, multiple 32. The installed selector uses 1024² pixels per MP and rounds dimensions to that multiple: the result is 640 × 640. Node #81 is 1.0 MP at 1:1, producing 1024 × 1024. The saved “768 × 768” titles and the disconnected 1344 × 768 / 1280 × 704 widget values do not describe the connected output dimensions.
Use a simple entrance or one pose, and shorten the timeline to match the duration control. Confirm identities, references, output and audio. Do not expect all four photographs from a five-second test.
Restore the 15-second target and full prompt. Watch for duplicate characters, changing glasses, floating hands, a second cat entrance, missing flashes and premature smiles. Keep the best performance, even if a timestamp lands slightly late.
Save a copy of the graph and record your seed. Fix one direction issue at a time. The supplied noise nodes randomize, so preserve their values when comparing prompt changes. Upscaling adds detail; it cannot reliably fix a broken performance.
Hardware demand depends on the exact quantized weights, memory/offload setup, canvas and duration. A graph title mentioning “8 GB” is not a tested minimum. No fresh GPU render or runtime benchmark was performed for this written guide.
08 / THE FIELD GUIDE
The photostrip is the scene’s emotional summary. It also makes a very good continuity check.

THE TAKEAWAY
Choose a clear moment immediately before or after each flash, when exposure is normal and the expression has settled. Stack the four images in story order with cream borders and even gaps.
Check the sequence as stills: do Kai’s glasses stay consistent, do Niko’s overalls remain teal, and does Mochi appear only when the story needs him? If the four photos read without the video, the scene has a clear visual arc.
Download the cute photostrip ↓Original 1152 × 4408 PNG. Selected from the finished clip, then arranged as a keepsake. No new poses were generated for this strip.
09 / THE FIELD GUIDE
Treat an imperfect result as information about what the model misunderstood.
Use ComfyUI Manager’s missing-node installer, restart ComfyUI, and refresh the page. Install KJNodes and MMH3 Ultimate Upscale as required. If a missing node is part of ComfyUI core, update the engine rather than searching for an unrelated node pack.
Use Manager’s model installer for listed models, or place the author’s matching file in the correct models subfolder. Re-select it in the loader. The graph does not bundle weights, and custom Wulver weights are not included in these downloads.
Keep one approved sheet per reference input. State that the panels are views of one individual, and render exactly one Kai, one Niko and one Mochi. Do not feed a collage of the entire scene as another identity reference.
Replace the entire old surprise and final-photo passages with the corrected prompt, not just a sentence. The entrance belongs in the third beat only. Explicitly keep Mochi hidden through the second flash.
Lock architecture and camera, not pedestrians. Ask background people to walk through the windows during the foreground pose holds. Keep their movement secondary.
Ask for completion, then a hold, then the flash. Reduce extra actions if necessary. Exact hand shapes are harder than broad silhouettes; choose the take with the clearest readable gesture.
Specify stepped key poses, sparse in-betweens and holds. Avoid optical-flow interpolation in the edit if it removes the intended cadence. Changing the export rate alone changes playback speed.
Test a shorter duration and a supported smaller canvas. Reduce other GPU workloads and review the model’s offload/quantization requirements. Keep a saved copy of the original graph; lower settings do not guarantee every GPU can run these weights.
Use the original download, not a screenshot, social-media copy or recompressed image. It must retain the workflow metadata. Try opening the supplied JSON as a fallback, then resolve dependencies and relink inputs.
MAKE THE NEXT ORDINARY MOMENT.
Keep the cast. Change the place. Write one small surprise worth watching.
Get the files and make your scene ↑