What Data Leaves Your Machine During Hosted Image-to-Video Generation
Hosted image-to-video services retain source images, generated clips, and session data under terms that vary widely—local generation is the only path where nothing leaves your machine.
What Data Leaves Your Machine During Hosted Image-to-Video Generation
Any source image you upload to a hosted platform is a copy you have given them permission to keep, unless their terms of service explicitly say otherwise. The privacy failure mode is direct. A source image uploaded to a platform with a training-data clause may become part of a future model's training set. You cannot undo that by deleting the output clip. The image is already ingested into their pipeline, logged, and potentially stored for content moderation review, human review, or model improvement. The only way to guarantee no data leaves your machine is to run the model locally. That means ComfyUI offline mode, a local GPU, and open-weights models. Everything else is a trade-off between convenience and control.
Hosted Image-to-Video API Data Retention: What Each Service Actually Keeps
Every hosted image-to-video platform retains data from your generation session. The differences are in what they retain, for how long, and what their terms of service permit them to do with it. Below is what the four most commonly used hosted platforms state in their public policies. Each claim is dated and attributed to a specific document section.
Runway Gen-3 Alpha Terms of Service Data Use
Runway's Terms of Service (section 5.2, dated 2025-10) state that Runway may use user-uploaded content, including source images and generated clips, for "product improvement, research, and development of new features." The document does not define a retention period for source images. The Acceptable Use Policy (2025-10) permits human review of generated content for safety purposes. The training-data inclusion clause is explicit: "Content you upload may be used to train and improve our models." There is no opt-out provision for paid tiers. Commercial use is tier-gated: the Free tier is non-commercial only; paid tiers permit commercial use but do not change the data retention or training-data terms.
Runway Gen-3 Alpha Source Image Retention
Runway's privacy policy (section 2.1, dated 2025-10) states that source images are retained for the duration of the user's account and for 30 days after account deletion, after which they are deleted from active storage. The policy does not address whether images remain in backups or training datasets after deletion.
Runway Gen-3 Alpha Generated Clip Storage
Generated clips are stored indefinitely on Runway's servers unless you delete them. The Terms of Service (section 5.3) grant Runway a license to display generated clips in portfolios, galleries, and promotional materials. You cannot prevent display.
Runway Gen-3 Alpha Human Review Access
The Acceptable Use Policy (2025-10) confirms that Runway conducts human moderation reviews of generated content for compliance with content policies. The policy does not specify whether reviewers see the source image or only the output clip.
Runway Gen-3 Alpha Training Data Inclusion
Section 5.2 of the Terms of Service (2025-10) explicitly permits use of uploaded content for model training. This clause applies to all users, including paid subscribers.
Luma Dream Machine Terms of Service Data Use
Luma's Terms of Service (section 4.3, dated 2025-11) state that user-uploaded content, including source images and generated clips, may be used for "training and improving machine learning models." The privacy policy (section 3.1, dated 2025-11) states that source images are retained for the duration of the account and for 90 days after deletion. Generated clips are stored until you delete them. Human review is permitted under the Content Policy (2025-11) for moderation purposes. Commercial use is tier-gated: the Free tier is non-commercial; paid tiers permit commercial use but do not alter data retention terms.
Luma Dream Machine Source Image Retention
Retained for account duration plus 90 days post-deletion. No specification on backup retention.
Luma Dream Machine Generated Clip Storage
Indefinite unless you delete them. Luma may display clips in promotional materials under section 4.4 of the Terms of Service.
Luma Dream Machine Human Review Access
Permitted for content moderation under the Content Policy. Reviewers may see both source and output.
Luma Dream Machine Training Data Inclusion
Permitted under section 4.3 of the Terms of Service. No opt-out.
Pika 2.0 Terms of Service Data Use and Google Veo 2 API Privacy Policy
Pika 2.0 Terms of Service Data Use
Pika's Terms of Service (section 6.1, dated 2025-10) state that uploaded content may be used for "model training, research, and service improvement." The privacy policy (section 2.2, dated 2025-10) states source images are retained for 60 days after account deletion. Generated clips are stored indefinitely. Human review is permitted under the Content Policy (2025-10) for moderation. Commercial use is tier-gated. Training-data inclusion is permitted with no opt-out.
Google Veo 2 API Privacy Policy
Veo 2 is accessed through Google Cloud Vertex AI. Google Cloud's Data Processing Agreement (dated 2025-12) states that source images and generated clips are retained for 30 days for API request logging and troubleshooting. Google's AI Prohibited Use Policy (2025-12) permits human review for safety. The Terms of Service (section 7.1) state that Google does not use customer content to train its models unless the customer explicitly opts into the Research & Model Improvement Program. Opt-out is available by default. Commercial use is permitted under standard Google Cloud terms. Retention after account termination: 30 days for API logs; generated clips are deleted upon manual deletion or account closure.
Google Veo 2 API Source Image Retention
30 days for API request logging. No indefinite storage for source images.
Google Veo 2 API Training Data Inclusion
Opt-in only. Default is no training data use. This is a significant difference from Runway, Luma, and Pika.
ComfyUI Local Generation Privacy: What Stays and What Leaks
ComfyUI local generation privacy is absolute if you run the model entirely offline. No data leaves your machine because there is no network request to a remote server. The source image is processed locally. The generated clip is stored on your local drive. No API request logging occurs. This is the only way to guarantee that your source image is not retained, reviewed by humans, or used for training-data inclusion.
The privacy failure mode for ComfyUI is node connections to external services. ComfyUI can load nodes that connect to cloud GPU rental services such as RunPod or Vast.ai. If you configure a node to offload inference to a remote GPU, the source image is transmitted to that server. The same data retention risks apply as with any hosted platform. Check the node documentation: if the node sends data to an external API, you are no longer running locally. The only safe configuration is offline mode with no external node connections and no internet access during generation.
Cloud GPU Rental Image-to-Video Privacy Policy
RunPod and Vast.ai are infrastructure providers; they do not access or retain the data you process on their GPUs. Their privacy policies (RunPod Privacy Policy, section 3.1, dated 2025-09; Vast.ai Privacy Policy, section 2.1, dated 2025-08) state that they do not inspect or store user data beyond the session duration. However, the model you load on the cloud GPU may itself make network calls. An open-weights model loaded from Hugging Face Diffusers does not inherently phone home, but a custom node or wrapper script might. The risk is not the GPU rental service; it is the software stack running on it.
What the EU AI Act Transparency Obligations Mean for Hosted Image-to-Video
The EU AI Act (Article 50, application date 2026-08-02) requires providers of AI systems that generate synthetic video to disclose that the output is artificially generated or manipulated. For any hosted platform operating in the EU, Runway, Luma, Pika, Google, OpenAI, this means the platform itself must label generated clips as AI-created. The disclosure is not optional for the provider. The exception is when the output is used for detection, prevention, investigation, or prosecution of criminal offences.
For the user, the obligation is different. If you publish a generated clip in the EU, you must not remove or obscure the provider's disclosure. The EU AI Act does not require you to add your own label, but platform policies (YouTube, Meta, TikTok) independently require you to mark altered or synthetic content. YouTube's policy (effective 2024-03-18) requires disclosure when content is altered or synthetic and appears realistic. Meta's policy (effective 2024-05-01) applies a "Made with AI" label when industry-standard signals are detected, including C2PA metadata. TikTok's policy (effective 2024-05-09) requires creators to label AI-generated content showing realistic scenes.
The practical consequence: even if a platform does not embed a visible watermark, it may embed metadata such as C2PA version 2.1 (specification dated 2024-06). Runway, OpenAI Sora, and Google Veo are reported to adopt C2PA metadata. If you strip that metadata using FFmpeg before uploading to a platform, you violate the platform's policy and risk content removal. Metadata stripping is detectable by platforms that compare file hashes against generation-platform signatures.
| Service | Source Image Retention | Generated Clip Storage | Human Review | Training Data Inclusion | Commercial Use |
|---|---|---|---|---|---|
| Runway Gen-3 Alpha | Account + 30 days post-deletion | Indefinite; may be displayed | Permitted for moderation | Permitted; no opt-out | Tier-gated (Free: non-commercial) |
| Luma Dream Machine | Account + 90 days post-deletion | Indefinite; may be displayed | Permitted for moderation | Permitted; no opt-out | Tier-gated (Free: non-commercial) |
| Pika 2.0 | Account + 60 days post-deletion | Indefinite | Permitted for moderation | Permitted; no opt-out | Tier-gated (Free: non-commercial) |
| Google Veo 2 API | 30 days for API logging | Until manual deletion or account closure | Permitted for safety | Opt-in only; default off | Permitted under Google Cloud terms |
ComfyUI Offline Mode and Open-Weights Models: What You Actually Give Up
Running locally with ComfyUI offline mode and open-weights models (Wan 2.1, LTX-Video 0.9, Stable Video Diffusion 1.1) eliminates all data retention risks. No source image leaves your machine. No API request logging occurs. No human review is possible. No training-data clause applies because there is no data to include.
What you give up: native resolution above 720p. Clip length of 2-5 seconds versus 5-10 seconds on hosted platforms. Motion coherence that is poorer than the leading hosted models because open-weights models do not have the same scale of training data. No camera-motion control beyond a single motion-strength slider for most open-weights models. No region-mask support. No start-and-end-frame support. Commercial use is restricted: Stable Video Diffusion 1.1 prohibits commercial use under the Stability AI Non-Commercial Research Community Licence; LTX-Video 0.9 prohibits commercial use under CC BY-NC 4.0; Wan 2.1 permits commercial use under Apache 2.0.
The VRAM requirement is real: Wan 2.1 recommends at least 16 GB VRAM; LTX-Video recommends at least 12 GB; Stable Video Diffusion recommends at least 16 GB. Peak VRAM usage during VAE decode and upscaling can exceed the stated figure by 2-4 GB. If your GPU cannot handle it, you crash mid-inference.
What Happens When You Strip Metadata
If you use FFmpeg to strip metadata from a locally generated clip, you remove the C2PA or other embedded signals. For local generation, there is no metadata to strip because the model does not embed any. The clip is clean. For hosted generation, stripping metadata before uploading to a platform violates the platform's synthetic-content disclosure policy. YouTube, Meta, and TikTok all require that the metadata remain intact or that you self-disclose. FFmpeg command: ffmpeg -i input.mp4 -map_metadata -1 -c:v copy -c:a copy output.mp4. This removes all metadata. Doing this to a hosted-platform clip is a policy violation.
The Single Most Practical Thing to Do Next
Before you upload a source image to any hosted image-to-video platform, read the specific terms of service section on data use and training-data inclusion. If the clause permits training data use and you cannot accept that risk, do not upload. Use ComfyUI offline mode with an open-weights model instead. The trade-off is lower resolution and shorter clips, but your source image never leaves your machine.
Common Questions
Can a hosted image-to-video service use my source image to train its next model?
Yes, unless the platform explicitly states otherwise. Runway, Luma, and Pika all include training-data clauses in their terms of service. Google Veo 2 is the exception: training data use is opt-in. Check the specific section of the terms you agreed to.
Does deleting my account remove my source image from a hosted service's training dataset?
No. The terms of service for Runway, Luma, and Pika state that source images are retained for 30-90 days after account deletion from active storage. They do not guarantee removal from training datasets or backups.
Is ComfyUI offline mode truly private?
Yes, if you run it entirely offline with no internet access during generation. The failure mode is misconfigured nodes that connect to external services like RunPod or Vast.ai. Check every node's documentation for network calls.
Does the EU AI Act require me to label my generated video?
The EU AI Act Article 50 (effective 2026-08-02) requires the provider to label the output. If you publish the clip in the EU, you must not remove that label. Platform policies (YouTube, Meta, TikTok) add separate requirements to self-disclose.