Node-based AI workflows. Local execution.
Same engine OpenArt runs in their cloud. Hybrig runs it on your rig. ComfyUI under the hood — eight production-grade graphs already wired and shipping in the desktop app, every one of them executing on the GPU you already paid for.
What the cloud-only crowd charges
Magnific charges ~$39 – $199/month {{TODO: source — magnific.ai/pricing}} to run an AI upscaler in their cloud. OpenArt charges per-credit on ComfyUI workflow execution {{TODO: source — openart.ai/pricing}}, so longer graphs and bigger images cost more every time you hit run. Hybrig runs the same node graphs on the GPU you already paid for. The bill is your power company.
Click a template, land in Studio with it pre-loaded.
Each template here is a real multi-node graph wired end-to-end. One click opens the Studio canvas with the nodes and edges already on the board, ready to fill in your inputs and hit Run.
Face Swap + Voice Dub
Face SwapDrop a video clip, your face, and a voice sample — get back the same scene with your face and voice.
8 nodes · 8 edgesOpen in Studio →Talking Head from Script
AvatarType a script, plug in your face and voice — get a video of you saying it.
8 nodes · 9 edgesOpen in Studio →Long-audio lipsync (chop + wrap + Seedance + stitch)
LipsyncDrop a long VO + a face. Hybrig chops at silences, wraps each segment, runs Seedance one at a time, and stitches the lot back into one continuous video.
7 nodes · 7 edgesOpen in Studio →Image to Video
VideoDrop a still photo, get a short animated video.
4 nodes · 3 edgesOpen in Studio →Character Replacement Video
Face SwapReplace the character in an existing clip with a reference face using local Wan VACE.
5 nodes · 4 edgesOpen in Studio →Voice Clone
VoiceProvide a voice sample and a script — get audio of that voice reading your script.
3 nodes · 2 edgesOpen in Studio →Lipsync to New Audio
LipsyncReplace what someone is saying in a video — keep their face, swap the dialogue.
4 nodes · 4 edgesOpen in Studio →Headshot Stills (Flux + LoRA)
ImageType a description, get a batch of stills using your trained LoRA.
2 nodes · 1 edgesOpen in Studio →Idea → Headshot
ImageType a rough idea — get a polished still using your LoRA.
3 nodes · 2 edgesOpen in Studio →Idea → Animated Image
VideoType a rough idea — get a still that animates into a short video.
4 nodes · 4 edgesOpen in Studio →Idea → Talking Head
AvatarType a topic — get a video of you talking about it in your voice.
9 nodes · 11 edgesOpen in Studio →Video Face Swap (your LoRA)
Face SwapReplace someone's face in a video with your trained LoRA — works frame-by-frame, requires a trained LoRA of you.
8 nodes · 9 edgesOpen in Studio →Prompt Cleanup
MarketingPaste any rough prompt — get a polished version with your LoRA trigger and cinematic descriptors.
2 nodes · 1 edgesOpen in Studio →Cohort Ad Pipeline (N prospects, 1 click)
CohortUpload a prospect CSV + Deep Intel + voice ref. Get one personalized ad per row, voice-cloned and leak-stripped. The pipeline behind Roof Radar's Peak Custom render.
8 nodes · 8 edgesOpen in Studio →Reference Plate Farm
FarmPlug in a character pack and shot list, then batch-generate clean reference plates for any storyboard or client ad.
7 nodes · 8 edgesOpen in Studio →Finger-circle reveal
EffectsPhone in hand or gesture-circle that frames a real product UI screenshot inside a live-action plate. Cloud polish on the circle interior is optional.
5 nodes · 4 edgesOpen in Studio →Background seasons hypercut
EffectsLock the subject, swap the backgrounds. One source plate, N season/weather variations, hard-cut on a beat.
5 nodes · 4 edgesOpen in Studio →Object additive buildup
EffectsOne locked plate, an object materializes in compressed time. Construction, growth, assembly — the lifespan in seconds.
5 nodes · 4 edgesOpen in Studio →Object swap replacement
EffectsSame plate, same lighting, swap one object for another. Before/after pair on a single held shot — clothing, vehicles, products, set dressing.
5 nodes · 4 edgesOpen in Studio →Background territory zoom-out
EffectsStart tight on a subject, pull back until a wider context surface fills the frame. Generated expansion, not filmed.
5 nodes · 4 edgesOpen in Studio →Object removal layer reveal
EffectsPeel a surface to expose what is underneath. Held plate, masked region, removal pass — the companion to additive-buildup.
5 nodes · 4 edgesOpen in Studio →Wordmark end-sting (music beat)
EffectsClosing shot of an ad. A surface in the world carries the wordmark; the music sting lands on the reveal. The last frame is the trophy.
6 nodes · 5 edgesOpen in Studio →
Eight workflows. Seven local. One cloud rescue valve.
Each card below is a real ComfyUI graph the desktop app builds at runtime. Node counts come from the live builder code — not a brochure. Local-by-default; the one cloud entry is the fallback path Hybrig takes only when the local chain misses on a critical shot.
Wan 2.2 5B Turbo (image → video)
Runs locallyAlibaba's Wan 2.2 TI2V 5B Turbo distill. Takes a still photo and a motion prompt, renders a 5–10 second clip. Default video model on Hybrig — runs on a 12 GB GPU, ~3–5 min per clip on a 4090.
11 nodes·ComfyUI graphBest for- Spokesperson clips and product demos on your own GPU
- High-iteration drafting where re-takes are free
- Single source-image workflows (UGC ad, founder shot, character motion)
Wan 2.1 14B (image → video, high quality)
Runs locallyThe full-fat Wan 2.1 image-to-video graph. 14B parameter diffusion + CLIP-Vision conditioning + 96-channel I2V encode. Slower than Turbo (~6–10 min per clip on a 4090) but tighter motion fidelity on demanding shots.
12 nodes·ComfyUI graphBest for- Clips where Wan 2.2 Turbo's 4-step distill drops fine motion detail
- Locked-off identity shots needing the full sampler recipe (30 steps, cfg 6)
- Fallback when 2.2 chokes on a specific install
Wan TI2V (text-only → video)
Runs locallyWan 2.2 5B's text-only branch. No source image. Type a prompt, get a 5-second clip from scratch. Same Wan stack minus the LoadImage and encode steps.
8 nodes·ComfyUI graphBest for- Mood shots, b-roll, scene transitions where you don't have a still yet
- Concept exploration before committing to a real photoshoot
- Prompt iteration on motion direction
Flux + LoRA (identity-trained character)
Runs locallyFlux dev with a trained per-character LoRA. The strongest still-image identity lock on Hybrig. Train once on 15–30 reference photos at /characters, then render the same face into any scene, wardrobe, or lighting forever.
11 nodes·ComfyUI graphBest for- Spokesperson stills, brand portraits, recurring character renders
- Multi-shot consistency across a campaign or series
- Prompt-driven scenes — LoRA holds identity, the prompt steers everything else
PuLID-Flux (face swap from one photo)
Runs locallySingle-photo face-identity conditioning for Flux via InsightFace embeddings + EVA-CLIP. No LoRA training required — drop in one reference photo, get the same face in a new scene. ComfyUI-PuLID-Flux-Enhanced custom node under the hood.
13 nodes·ComfyUI graphBest for- Putting a face into a new scene without training a LoRA
- Quick identity conditioning when you only have one good reference photo
- Style swaps where the face stays locked but the prompt rewrites everything else
EchoMimicV2 (audio-driven talking head)
Runs locallyDrives a still portrait with an audio file — mouth and head move to match the speech. Trained on Echo's audio→motion mapping. Six nodes: load image, load audio, load model, predata, sampler, save video.
6 nodes·ComfyUI graphBest for- Lipsync from a still spokesperson photo + a recorded script
- Bring-your-own-audio workflows (record yourself, drive the avatar)
- Multilingual VO over the same face
LatentSync (alternative lipsync)
Runs locallyDrives an existing video clip's mouth movement to match a new audio track. Lighter graph than EchoMimic — four nodes — and feeds on a video instead of a still. Good for re-dubbing existing footage.
4 nodes·ComfyUI graphBest for- Re-dubbing existing video clips with new VO or translated audio
- Tighter mouth sync on shots where the body motion is already correct
- When EchoMimic's still-driven path can't preserve the original camera move
Seedance fallback (cloud rescue valve)
Cloud fallbackByteDance Seedance 2.0 via fal.ai. Cloud-only, billed per second of output. NOT the default — Hybrig reaches for this only when the local chain (Wan 2.2 → Wan 2.1 → Hunyuan → LTX) misses on a critical shot. Top-tier face lock under heavy motion + native audio mux. For lipsync calls, Hybrig wraps the audio into a blank-visual video container before handing it to Seedance — raw audio in causes the model to reshape the tempo and hallucinate lyrics, and two creators independently surfaced the workaround. We bake it in so you never trip on it.
Cloud endpoint·fal.aiBest for- Premium brand finals where face lock under motion is non-negotiable
- Cinematic VFX and lighting consistency cloud models still own
- When you've already tried local and the shot needs the rescue valve
Gemini Omni polish (cloud — hero beats only)
Cloud fallbackGoogle DeepMind's Gemini Omni via Google AI subscription. Conversational video editing, physics-aware motion, world-knowledge compositing. NOT a farm engine — per-render meter + subscription gate break cohort economics the moment you scale. Hybrig wires Omni as a polish layer for the one or two hero beats per spot where physics simulation (water, gravity, kinetic objects) or world-knowledge context (a specific historical scene, cultural reference) gives a shot something local Wan cannot fake. You finish your local farm first, identify which beats genuinely need it, then send only those to Omni.
Cloud endpoint·fal.aiBest for- Hero beats with physics simulation local Wan can't fake (water, gravity, kinetic objects)
- Compositing that needs world-knowledge context (historical scene, cultural reference)
- Iterative conversational refinement on a SINGLE finished beat — not the whole farm
Same graphs. Different bill.
OpenArt is a ComfyUI workflow gallery that runs the graphs in their cloud and charges per execution. Magnific is a single-purpose AI upscaler that charges a flat monthly subscription regardless of usage. Hybrig is the third option: same engine, your GPU.
Run any of these workflows on your rig
Install the desktop app. It bundles the local model supervisor, points the node-graph builder at your GPU, and signs you in to your Hybrig account. Every workflow on this page is wired and ready — Wan, Flux, PuLID, EchoMimic, LatentSync — running on the GPU you already paid for, no per-credit meter, no monthly cap.