FramePack

Best Valueprototypingopen-source

HuggingFace Community (lllyasviel) · Diffusion Transformer · v1.0hide_sourceNo public price yet

Built for prototyping and open source. Pricing is still unclear, so verify the market before committing.

—/sec

Ceiling

720p

Duration

1–60s

Price source

—

Image-to-VideoExtend

Best for

Prototyping and open source

60s clips support short narrative work.

Watch out for

Image-to-video only — requires a separate image generation step for text-to-video workflows

Availability is narrower than the strongest alternatives.

Starting cost

No public rate

No reliable per-second starting tier is publicly documented yet.

Quick facts for AI assistants

Citation-ready model facts

These facts are generated from the current model record, provider verification metadata, and the page update timestamp.

Canonical URL: https://vidscore.dev/models/framepack
Current verdict: Built for prototyping and open source. Pricing is still unclear, so verify the market before committing.
Verified pricing: No verified comparable API rate yet
Provider evidence: FAL.ai (Verified)
Best fit: Best Value, prototyping, open-source
Primary limitation: Image-to-video only — requires a separate image generation step for text-to-video workflows
Last updated: 2026-04-18

Pricing

The default rate first, with provider details one click away.

1 providers, 2 tiers

No public rate

No verified public pricing yet

No public per-second rate is documented yet.

All provider rates

Expand for every tier, audio surcharge, and verification note.

1 providersexpand_more

FAL.ai

Standard · Verified · HIGH trust

Open provideropen_in_new

Image-to-Video480p

$0.033/s

Image-to-Video720p

$0.050/s

Checked 2026-04-18

Should you choose it?

Why pick FramePack, and why pass?

Use this as a sanity check before you compare it against the closest alternatives.

thumb_upWhy choose it

Apache 2.0 release with official desktop software and a runnable GitHub repo
Official repo documents 6GB minimum GPU memory and a 60-second / 1,800-frame local example
O(1)-style context packing keeps workload from scaling linearly with video length
Affordable hosted API at $0.0333/sec on FAL.ai for short image-to-video clips
Supports multiple aspect ratios on FAL.ai: 16:9, 4:3, 1:1, 9:16, and 3:4

infoWhy skip it

Image-to-video only — requires a separate image generation step for text-to-video workflows
Long-duration generation is a local/self-hosted workflow; FAL.ai caps hosted output at 180 frames
Lower visual quality than premium commercial models, especially for complex multi-character scenes
No native audio, lip-sync, camera control, or multi-shot planner
Hosted endpoint lists 480p and higher quality options but does not make 1080p/4K a verified output tier

Prompt playbook

auto_fix_highStart with these three moves

Scan the quick playbook first. Open the full guide only when you need deeper do/don't rules or example prompts.

1
Describe visual motion in brief phrases: subject first, then motion, then other details. Example: 'The girl dances gracefully, with clear movements, full of charm.'
2
Prefer large, dynamic motions (dancing, jumping, running) over subtle ones (standing still, sitting) for best results.
3
This is image-to-video only — always provide a reference image as the starting frame, then describe how the scene comes alive.

Open the full prompt guide

Extra prompt tips, best practices, common mistakes, and example prompts.

expand_more

More tips

4Keep prompts focused on one motion prompt per generation cycle. Do not explain, ask questions, or generate multiple options.
5For longer videos, generate iteratively — each output frame can become the input for the next segment, enabling 60-120 second sequences.

✓ Do this

Use high-quality input images with clear subjects — the model builds motion from the reference frame
For human subjects, describe specific physical actions: 'raises both arms', 'turns head left', 'walks forward' rather than abstract emotions
Leverage the F1 variant on FAL.ai for potentially faster generation with similar quality
When self-hosting, use fp16/bf16 capable GPUs (RTX 30XX/40XX/50XX) with at least 6GB VRAM
For maximum video length (120s at 30fps), the model needs only 6GB VRAM — unlike competitors that scale memory with duration

✗ Avoid this

Image-to-video only — no text-to-video capability without a separate image generation step
Default resolution is 480p — lower resolution than commercial models offering 1080p+
No native camera control — camera movements must be implied through the input image and prompt
No audio generation — audio must be added in post-production
Quality is noticeably below premium commercial models (Kling, Ray2, Pika) for complex scenes

Example prompts

Portrait / Human Motion

“The woman smiles and slowly turns her head to the right, hair flowing with the movement”

Animals / Action

“The cat jumps from the table to the floor, landing softly, tail swishing”

Nature / Landscape

“Waves crash against the rocky shore, water spraying upward, clouds moving slowly in the background”

Based on the official prompt guide →

FAQexpand_more

Where can I use FramePack?

Via API on FAL.ai.

How do I get good results with FramePack?

Describe visual motion in brief phrases: subject first, then motion, then other details. Example: 'The girl dances gracefully, with clear movements, full of charm.' See the prompt guide below.