Guaardvark v2.5.1 makes Wan2.2 the headline video pipeline. Generate 5–30 second clips from text prompts or animate still images, all through a ComfyUI backend running entirely on your own GPU. CogVideoX is still supported for faster iteration. No cloud APIs, no per-second billing, no usage caps — just your RTX and a prompt.
Wan2.2 is the new headline pipeline for high-fidelity 5–30 second clips. CogVideoX remains available for faster generation and rapid prompt iteration. Both support text-to-video (t2v) and image-to-video (i2v) modes with configurable resolution, step count, and guidance scale.
Every inference step runs on your own NVIDIA GPU via ComfyUI. Requires 12 GB+ VRAM; an RTX 4070 Ti SUPER generates a 5-second Wan2.2 clip in approximately 4–7 minutes at 480p. Your prompts, source images, and finished videos never leave your machine.
Submit multiple generation jobs as Celery background tasks. Walk away while clips render overnight. Each job carries its full parameter set — prompt, model, resolution, frame count, post-processing flags — and results land in the Content Library with full metadata for comparison and re-use.
Generate at 480p for speed, then feed each frame through Real-ESRGAN in the Upscaling pipeline for 1080p+ final output. RIFE frame interpolation smooths the base frame rate before upscaling. The combined pipeline delivers professional-quality output without re-running the expensive diffusion step.
Guaardvark’s video generation pipeline is built on ComfyUI running as a managed backend service. ComfyUI handles the generation graph construction, noise scheduling, temporal attention, and VAE decoding for both Wan2.2 and CogVideoX. From the Guaardvark interface, you write a natural-language prompt, choose your model, set resolution and frame count, and submit. The platform translates your parameters into a ComfyUI workflow JSON, queues it as a Celery task, and returns a job ID. The ComfyUI worker executes the graph on your GPU and deposits the finished video file into data/outputs/. You can monitor progress, inspect intermediate frames, and queue additional jobs without leaving the interface — or without the interface open at all, since jobs run as background workers that persist across browser sessions.
The image-to-video path works by routing a conditioning image through Wan2.2’s or CogVideoX’s image encoder before the diffusion process begins. The encoder preserves the visual identity of the source image — composition, colour palette, subject matter — while the temporal attention layers generate plausible motion dynamics. The result is a video that brings your still image to life in a way that respects the original rather than diverging into a new scene. This is useful for animating product shots, concept art, storyboard panels, or any AI-generated image that benefits from implied motion. An optional text prompt can guide the direction of motion without overriding the visual identity of the source.
Clip lengths range from 5 to 30 seconds. Shorter clips (5–8 seconds) are well-suited to social media loops, product previews, and animated illustrations. Longer clips (15–30 seconds) work for explainer intros, scene-setting sequences, and any content where sustained motion matters. For longer final pieces, generate multiple clips and stitch them in the Video Editor. Resolution is configurable; 480p is the recommended starting point for most consumer GPUs, with Real-ESRGAN upscaling delivering 1080p+ quality in the post-processing stage without re-running the full diffusion pass.
ComfyUI integration. Guaardvark manages a ComfyUI server process as part of its service stack. The integration layer in backend/services/video_generation_service.py constructs ComfyUI workflow JSON from the user’s parameters, POSTs it to ComfyUI’s /prompt endpoint, and polls the /history endpoint for completion. Model weights for Wan2.2 and CogVideoX are placed in ComfyUI’s models/ directory; the GPU Plugin Manager monitors VRAM before each job and can evict idle LLM or image-generation models to make room if needed. Wan2.2 requires 12 GB+ VRAM for base operation; CogVideoX fits in 10 GB with attention slicing enabled. Both models are downloaded via the standard model management UI in Settings → Models.
Post-processing pipeline. After ComfyUI returns the raw video frames, Guaardvark’s post-processing chain runs in sequence: first RIFE frame interpolation (if enabled) doubles or quadruples the effective frame rate by synthesising intermediate frames from optical flow estimates; then Real-ESRGAN super-resolution (if enabled) upscales each frame to the target output resolution. Both stages run as additional Celery tasks chained to the generation job, so they appear as sub-steps in the job progress panel. The final output is assembled into an MP4 container by ffmpeg and registered in the Content Library. All intermediate frames are kept in a job-specific temporary directory and cleaned up automatically after the final MP4 is confirmed.
# Example prompts and expected runtimes on RTX 4070 Ti SUPER (16 GB VRAM)
# Wan2.2 t2v, 480p, 5 seconds, 20 steps:
"A lone lighthouse on a rocky coast at sunset, waves crashing, cinematic"
# → ~4–5 min generation + ~1 min RIFE + ~3 min ESRGAN upscale to 1080p
# Wan2.2 i2v, 480p, 8 seconds, 20 steps (source: product-shot.png):
"Subtle product rotation, studio lighting, clean background"
# → ~7–9 min generation + post-processing
# CogVideoX t2v, 480p, 6 seconds, 50 steps (faster iteration):
"Abstract geometric shapes morphing in neon colours"
# → ~3–4 min generation
# Batch queue via llx CLI:
llx generate video "prompt one" --model wan2.2 --seconds 5
llx generate video "prompt two" --model wan2.2 --seconds 8
llx generate video "prompt three" --model cogvideox --seconds 6Generate dozens of 5–10 second clips for reels, shorts, and looping background videos overnight as a batch queue job. Because your prompts and output never touch a cloud service, you can create content around proprietary products, unreleased branding, or sensitive campaigns without any risk of content leakage or API terms-of-service violations. Pair with Video Text Overlay to burn captions automatically for silent-autoplay compliance.
Use the image-to-video path to bring still product photography or AI-generated concept art to life. Supply a source image and a short motion prompt (“subtle product rotation, studio lighting”), let Wan2.2 animate it, and run the result through Real-ESRGAN upscaling to 1080p. The output is suitable for e-commerce listings, pitch decks, and portfolio pieces — produced entirely in-house, with no creative brief shared with a third-party service.
Combine multiple Guaardvark features into a complete in-house production workflow: write the script in Documents, generate narration with Audio Foundry’s Piper TTS, produce scene clips with Wan2.2, stitch them with the Video Editor, burn captions via Video Text Overlay, and upscale the final cut to 1080p. Every step runs on your hardware. The result is a polished explainer video that touched no cloud service from script to export.
Commercial text-to-video services — Sora, Runway, Kling, Luma — charge per second of generated video, typically $0.05–$0.30 per second depending on resolution and model tier. A batch of 50 five-second clips at $0.10/s costs $25 — every time you run it. They also route your prompts and any conditioning images through their servers, making them unsuitable for proprietary content. Guaardvark’s video generation has zero per-generation cost after model download: the only expense is electricity. An RTX 4070 Ti SUPER draws roughly 250 W during inference; generating 50 five-second Wan2.2 clips takes approximately four GPU-hours, costing pennies in electricity. Your prompts, your source images, and your output videos never leave your network. There are no content filters, no usage quotas, and no terms-of-service restrictions on what you generate — you are constrained only by your own GPU and your own creative choices.
Wan2.2 requires a minimum of 12 GB VRAM for the base text-to-video pipeline at 480p. An RTX 4070 Ti SUPER (16 GB) is the development reference GPU; a 5-second clip at 480p takes 4–7 minutes at 20 inference steps. RTX 3090 and 4090 are also well-tested. RTX cards with 8 GB VRAM can run CogVideoX with attention slicing and tiled VAE decoding, but not Wan2.2 at default settings. AMD GPU support via ROCm is experimental. CPU inference is not recommended for video generation due to prohibitive generation times.
Wan2.2 is the headline pipeline in v2.5.1: it delivers higher visual fidelity, better temporal coherence across longer clips, and stronger adherence to complex scene descriptions. It requires more VRAM and generates more slowly. CogVideoX has a more compact architecture that generates clips faster, making it well-suited for rapid prompt exploration and scenarios where turnaround time matters more than peak quality. Both support text-to-video and image-to-video modes. CogVideoX is not being deprecated — it remains a fully supported option and is recommended for iteration before committing to a Wan2.2 final render.
Clip length is configurable from 5 to 30 seconds. Longer clips require more VRAM and more inference time in proportion to length: a 30-second clip takes roughly six times as long as a 5-second clip at the same resolution and step count. For content longer than 30 seconds, the recommended approach is to generate multiple shorter clips with consistent style prompts and stitch them together in the Video Editor.
Yes. The post-processing pipeline includes Real-ESRGAN super-resolution, which upscales each frame from the generation resolution (typically 480p) to your target output resolution (1080p or higher). This runs as a separate Celery task chained to the generation job. You can also run RIFE frame interpolation before upscaling to smooth the base frame rate. The combined RIFE + ESRGAN pipeline adds approximately 4–6 minutes to a 5-second clip on an RTX 4070 Ti SUPER, but the quality improvement is significant. Upscaling also pairs with the dedicated Upscaling feature page for standalone frame upscaling workflows.
Yes. Video generation jobs run as Celery background workers, so you can continue using chat, RAG search, image generation, and other features while clips render. The GPU Plugin Manager monitors VRAM usage across all active tasks. If a foreground task (such as an LLM inference call) needs GPU resources urgently, the manager can pause or throttle a background video job temporarily and resume it automatically when resources are available. Priority rules are configurable in Settings → GPU.