Wan 2.1 · Text to video
ConfiguredWan 2.1 T2V 1.3B
The CPU-first workhorse. A Q6_K GGUF denoiser with four-step distilled sampling for fast, repeatable text-to-video.
A private, CPU-only AI video studio. Text-to-video and image-to-video with a real rendering queue, measured status and MP4 downloads — no rented GPUs required.
Decorative preview · not a generated sample
The monitor above is interface art. Promptora never displays a synthetic clip as a model result.
CPU inference
No CUDA / ROCm
Model targets
Wan 2.1 + Wan 2.2
Queue
Persistent · SQLite
Privacy
Local files only
The studio
No black box, no invented progress. Every stage reports what it is doing, and the planner tells you when the hardware is not enough.
Jobs live in SQLite, not in memory. The worker recovers after a crash, closes stale sessions and resumes watching the queue.
The plan is a refusal surface, not a sales pitch. If RAM, resolution or model capability does not fit, you see the reason before a job is queued.
Opaque server-side sessions in HttpOnly cookies. Uploads are decoded, metadata-stripped and re-encoded as PNG.
Every completed render is ffprobe-verified, thumbnailed and served with HTTP range support for scrubbing and download.
Text-to-video for pure prompts. Image-to-video with JPG, PNG or WebP input that is validated by content, not filename.
Model shelf
Two production targets, one clearly quarantined experiment, and an optional diagnostic that never pretends to be AI.
Wan 2.1 · Text to video
ConfiguredThe CPU-first workhorse. A Q6_K GGUF denoiser with four-step distilled sampling for fast, repeatable text-to-video.
Wan 2.2 · Text + image to video
ConfiguredOne 5B target for both modes. Q4_K_M weights and a Turbo LoRA keep the memory footprint inside a CPU host.
LTX · Experimental
ExperimentalVisible for transparency, not for jobs. Upstream verification covered CPU denoise only; decode and encode are unproven.
From prompt to file
Promptora keeps generation and serving separate, so the studio stays responsive while a CPU render is running.
Write a prompt, choose a model and frame shape, then attach an image for I2V. Nothing leaves your server.
The planner resolves variants, checks peak RAM and frame limits, and refuses honestly when the host cannot run it.
A separate worker owns the queue. It downloads weights once, denoises, decodes, encodes H.264 and probes the MP4.
Completed clips land in My Creations with a thumbnail, download, retry and regenerate actions.
Wan 2.1 T2V 1.3B and Wan 2.2 TI2V 5B are the supported targets. Their readiness is checked against this host's available memory — a model can be installed and still be refused if it would exhaust RAM. LTX-Video is visible for transparency but disabled. The pipeline self-test produces a labelled test card, not AI footage.
Inspect live model readinessThe first frame is yours
Open the studio, write a shot and let the planner tell you the truth about this machine.