> For the complete documentation index, see [llms.txt](https://help.intangible.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://help.intangible.ai/visualize/ai-models.md). # Models Every image and video model the visualizer exposes, with one-line guidance on when to reach for each. The visualizer is a model aggregator: under the hood, every Generate Image and Generate Video click routes to one of several diffusion models from external providers. The dropdown is the working set, refreshed as providers ship new versions and as some get retired. This page is the working reference. Pick a model based on what your shot needs; the table below has one-line guidance per model. ## Image models | Model | Provider | When to reach for it | | -------------------------- | ----------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **Flux2 Pro** | Black Forest Labs | Next-generation Flux. Cleaner subjects, finer detail, better at rendering text in the image when text matters. | | **Nano Banana Pro** | Google | The cinematic workhorse. Strongest with image references, especially for character consistency across shots. Default reach for hero shots when you have reference imagery. | | **Nano Banana 2** | Google | Sharper than Pro on certain stylized treatments. Picky about prompts; rewards specificity. | | **GPT Image 2** | OpenAI | OpenAI's image model in the working set. Different stylistic register from the Flux and Nano Banana families; useful as an alternative when the others flatten the subject. | | **GPT Image 2.5 Sunburst** | OpenAI | The slower, higher-reasoning tier of GPT Image 2.5. The faster Flare tier mismatched likenesses when two characters each carried their own reference sheet. | | **Seedream 5 Pro** | ByteDance | Holds composition harder than anything else in the working set. Reach for it when the shot you framed has to survive the render intact, and when detail named in the description needs to outrank the reference images on everything except identity. | ## Video models ![The video model dropdown, grouped by vendor, with the mode shown as Video (from keyframes)](https://1179478100-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FSWrjIqlgVAblx6vHGJiM%2Fuploads%2Fgit-blob-bb72968eb414127ff6572f1b36c7d96f4c127c18%2Fvideo-model-dropdown-01.png?alt=media) | Model | Provider | When to reach for it | | --------------------- | ----------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **Veo 3.1** | Google | Wide latitude, slow generation. Best for atmospheric long-take shots where motion is naturalistic and you have time to wait. | | **Veo 3.1 Fast** | Google | Faster Veo at lower fidelity. Iteration-tier for atmospheric shots. | | **Kling 2.5 Turbo** | Kuaishou | Fast iteration. Use when you're cycling through compositions and want feedback in seconds. | | **Kling 2.6 Pro** | Kuaishou | Action-shot strong. Reach for it on motion-heavy shots. | | **Kling 3 Standard** | Kuaishou | Latest Kling family standard tier. | | **Kling 3 Pro** | Kuaishou | Latest Kling family pro tier. | | **Kling 3 (4K)** | Kuaishou | The 4K tier of the Kling 3 family, on its own endpoint. Reach for it when the deliverable needs 4K out of the visualizer rather than an upscale afterwards. | | **Kling o3 Pro** | Kuaishou | Available for both image-to-video and video-to-video. Takes up to eight reference images named in the prompt. Reach for it when a video shot has to honor specific reference imagery. | | **Omni Flash** | Google | Now on version 1.1. Available for both image-to-video and video-to-video, with generated audio, up to eight reference images, and an optional end frame on the image-to-video path. | | **Luma Ray 3.14** | Luma | Available for both image-to-video and video-to-video. Honors authored 3D structure strictly, so it is the reach when your shot has animated meshes, splined cameras, or specific motion the model should follow rather than reinterpret. | | **Luma Ray 3.2** | Luma | Video-to-video model. Takes an existing video as input and reskins or extends it while respecting the source motion. Higher fidelity; use when output quality matters more than cost. | | **Runway Gen 4.5** | Runway | Alternative video model. Different tonal register than Kling and Veo; useful for stylistic variation. | | **Seedance 2.0** | ByteDance | Available for both image-to-video and video-to-video. Motion-strong with first-and-last-frame support and up to seven reference images. Strong for photorealistic, faithful output – a good reach for technical or aerospace work where the render needs to stay true to the source. | | **Seedance 2.5** | ByteDance | The long-clip Seedance, on both paths. Image-to-video runs up to 30 seconds and carries up to 28 reference images alongside first and last keyframes, at 480p or 720p only. | | **Seedance 2.0 Mini** | ByteDance | The fast, low-cost tier of Seedance 2.0, on both paths. Same seven reference images and first-and-last-frame support, at 480p or 720p only. Reach for it while you are still deciding the motion, then re-render on 2.0 once the shot is locked. | | **Wan 3.0** | Alibaba | Available for both image-to-video and video-to-video, and the one entry that spans 480p, 720p and 1080p. Image-to-video takes a first and last keyframe and runs up to 30 seconds; video-to-video restyles a source clip. Neither path exposes an aspect ratio: it is derived from the input media. | | **FLUX.3** | Black Forest Labs | Image-to-video only, and it takes no reference images. First and last keyframes, 720p or 1080p, clips up to 20 seconds. Reach for it when the shot is carried by its keyframes rather than by a reference strip. | {% hint style="info" %} **Tuning Luma for technical and aerospace work.** Luma exposes a creativity control. Set it to **Adhere** with the strength low when faithfulness to the source geometry matters – it keeps the render true to what you built. **Reimagine** gives the model far more license and will invent elements that aren't in your scene, so keep it away from technical work. For faithful output, Seedance 2.0 is also a strong choice. {% endhint %} {% hint style="info" %} **Some models run on both paths, some only on one.** Ray 3.14, Omni Flash, Kling o3 Pro, Seedance 2.0, Seedance 2.5, Seedance 2.0 Mini and Wan 3.0 each appear for image-to-video and for video-to-video. Ray 3.2 is video-to-video only; FLUX.3 is image-to-video only. Video-to-video takes an existing clip as the source input rather than generating from the scene, so use it to restyle, extend, or relight footage you already have. The mode and strength control (Adhere → Reimagine) behaves the same on either path: low Adhere keeps the motion faithful to the source; Reimagine lets the model depart significantly from it. {% endhint %} {% hint style="warning" %} **Video-to-video sources have a duration ceiling that varies by model.** Seedance 2.5 accepts up to 30 seconds, Ray 3.2 and Ray 3.14 up to 18, Seedance 2.0, Seedance 2.0 Mini, Kling o3 Pro and Wan 3.0 up to 15, Omni Flash up to 10. A source longer than the selected model's ceiling is blocked before any credits are spent, with a dialog naming the limit. {% endhint %} {% hint style="info" %} **480p is on three models only.** The video ladder runs 480p, 720p, 1080p, 4K, and no single model covers all four. Seedance 2.5, Seedance 2.0 Mini and Wan 3.0 are the three that offer 480p; the first two stop at 720p, so picking either for a 1080p deliverable means regenerating on another model. Wan 3.0 is the one entry that spans 480p to 1080p, and the 4K paths are Kling 3 (4K), Seedance 2.0 and Omni Flash. On Omni Flash, 1080p and 4K are upscales of the same generation rather than native renders. Cost scales with the resolution you pick – see [Resolution and cost](/visualize/resolution-and-cost.md). {% endhint %} ## Which AI model should I use? Direct answers per common job: | If your shot is | Reach for | Why | | -------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | Brand-true product hero with reference imagery | **Nano Banana Pro** | Strongest reference-image conditioning. Holds character and product across shots. | | Photoreal still with detail that has to hold up close | **Flux2 Pro** | Black Forest Labs' current still model. Finer detail on the subject, with reference support. | | In-image text rendering matters (signage, packaging) | **Flux2 Pro** | Cleanest text generation in the working set. | | Stylized illustration or graphic look | **Nano Banana 2** | Sharper on stylized treatments. Reward specificity in the prompt. | | Atmospheric long-take video, naturalistic motion | **Veo 3.1** | Wide latitude. Slow but the best for ambient mood. | | Action-heavy video with motion | **Kling 2.6 Pro** | Action-shot strong. The chase-and-vehicle workhorse. | | Authored 3D motion needs to be honored exactly | **Luma Ray 3.14** | Honors keyframed and splined motion most strictly. | | Composition has to survive the render intact | **Seedream 5 Pro** | Holds the framing you built harder than the rest of the image working set. | | First-and-last-frame video interpolation | **Veo 3.1**, **Kling 2.6 Pro** or later, **Seedance 2.0**, **Seedance 2.5**, **Seedance 2.0 Mini**, **Wan 3.0**, **FLUX.3**, or **Omni Flash** | These model families support end frames. See [First and last frame](/visualize/first-and-last-frame.md). | | Video shot that has to follow specific reference imagery | **Seedance 2.5**, **Omni Flash**, **Kling o3 Pro**, or **Seedance 2.0** | These carry reference images into video prompts. Seedance 2.5 takes the most, up to 28. See [Generate video](/visualize/generate-video.md). | | Video with generated audio | **Omni Flash** or **Veo 3.1** | Both generate a synchronized audio track. Audio adds cost on top of the render. | | 4K video straight out of the visualizer | **Kling 3 (4K)**, **Seedance 2.0** or **Omni Flash** | The three 4K paths in the working set. See [Resolution and cost](/visualize/resolution-and-cost.md). | | Video-to-video (restyle or extend an existing clip) | **Luma Ray 3.2**, **Luma Ray 3.14**, **Wan 3.0**, **Omni Flash**, **Kling o3 Pro**, **Seedance 2.0**, **Seedance 2.5**, or **Seedance 2.0 Mini** | Supply a source video; the model reskins or extends it. Ray 3.2 for final quality, Ray 3.14 for iteration, Seedance 2.5 for a source longer than 18 seconds. | | Fast iteration on a video composition | **Kling 2.5 Turbo** or **Veo 3.1 Fast** | Seconds-tier feedback. | | Cheap, fast pass on video motion | **Seedance 2.0 Mini** | The fast tier of Seedance 2.0 at 480p or 720p. Iterate here, commit on 2.0. | | Single take that has to run past 15 seconds | **Seedance 2.5**, **Wan 3.0**, or **FLUX.3** | Seedance 2.5 and Wan 3.0 offer image-to-video durations up to 30 seconds; FLUX.3 reaches 20. | | Image edit where the faster tier already failed | **GPT Image 2.5 Sunburst** | The slower, higher-reasoning tier of GPT Image 2.5. The faster Flare tier mismatched likenesses when two characters each carried their own reference sheet. Whether Sunburst holds them apart isn't established, so treat it as the next thing to try rather than a fix. | ## Nano Banana Pro vs Seedream 5 Pro The default image model against the one you pick on purpose. Both take reference images; they differ in what they're tuned to honor. | | Nano Banana Pro | Seedream 5 Pro | | ------------------- | ------------------------------------------------ | ------------------------------------------------------------------------------------- | | What it honors most | The reference images, across shots | The composition you framed | | Reference behavior | Strongest for character consistency shot to shot | Detail named in the description outranks the references on everything except identity | | Best for | Hero shot with character or product references | A shot whose framing has to come back intact | | Where it sits | The visualizer's default image model | A deliberate pick | For a campaign that needs multiple shots of the same product, Nano Banana Pro. For a single frame you composed carefully and need back as framed, Seedream 5 Pro. ## What to do past the first pass The answer is always "test against your specific shot". Different models honor different prompt languages differently; what reads cleanly on Nano Banana 2 may flatten on Flux2 Pro and vice versa. Burn 1K iterations comparing before committing video credits. ## When models change The model list churns. Providers ship new versions, sometimes retire older ones, occasionally pull capabilities out of free or lower tiers. The visualizer's dropdown is authoritative for what's currently available; this page tracks the names and the rough heuristics. If a model you've been using disappears from the dropdown, the most likely reason is the provider deprecated it. **You don't have to hunt for the replacement yourself.** When a saved shot points at a model the working set doesn't carry, regeneration routes to a successor – a configured one where the model has one, otherwise the default for that kind of job – and an in-app banner names the substitution before the render runs. The same banner appears when a team admin has restricted the model you picked. Read the banner: the shot you get back came from a different model than the one stored on the shot, so the look can shift. If it matters, pick the successor deliberately and re-render. Models the visualizer routes away from, and where they land: | Stored on the shot | Regenerates on | | ------------------ | ---------------- | | Kling 2.0 Master | Kling 3 Pro | | Kling 2.1 Standard | Kling 3 Standard | | Kling 2.1 Pro | Kling 3 Pro | | Kling 2.1 Master | Kling 3 Pro | | Luma Ray 3 | Luma Ray 3.14 | | Luma Ray 2 | Luma Ray 3.14 | | Luma Ray 2 Flash | Luma Ray 3.14 | | Flux Pro Kontext | Nano Banana Pro | ## Older names you may see in transcripts Some YouTube tutorials and webinar talks reference earlier model names. The mapping: * *Kling 2.6* in the older transcripts is now **Kling 2.6 Pro**. * *Kling 2.0* and *Kling 2.1* in either the Standard, Pro, or Master tier map to the **Kling 3** family. * *Luma Ray 3*, *Ray 2*, and *Ray 2 Flash* map to **Luma Ray 3.14**. * *Seedance 2.0 Motion Guide* is the video-to-video path of **Seedance 2.0**. * *Flux Pro Kontext* is retired and no longer in the dropdown. Shots stored on it regenerate on **Nano Banana Pro**. * *Flux 1 Depth* (referenced in summit talk) is superseded by **Flux2 Pro** for general use. * *Flex* (mentioned in passing) was an older naming; **Flux2 Pro** is the current Black Forest Labs image model in the working set. Use the current names everywhere in your work. The older names will still appear in older tutorial videos until the team re-records. ![Image model dropdown expanded with provider-grouped options: Black Forest Labs (Flux Pro Realtime, Fast Pro) and Google (Nano Banana Pro, Nano Banana 2)](https://1179478100-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FSWrjIqlgVAblx6vHGJiM%2Fuploads%2Fgit-blob-013e543561a20603432e189cb54c53a9b859e916%2Fimage-model-dropdown-01.png?alt=media) ## Watch {% embed url="" %} ## Provider documentation Each provider has its own deeper docs for what their models can and cannot do at the system-prompt level: * Black Forest Labs Flux: [bfl.ai](https://bfl.ai) * Google Veo: [aistudio.google.com/models/veo-3](https://aistudio.google.com/models/veo-3) * Kuaishou Kling: [kling.ai](https://kling.ai) * Luma Ray / Agents: [lumalabs.ai/ray](https://lumalabs.ai/ray) * Runway: [runwayml.com](https://runwayml.com) Most users won't need the provider docs; the visualizer abstracts away most provider-specific syntax. For specialized model behaviors not exposed in Intangible's UI, the provider docs are the next stop. ## Related * [Generate image](/visualize/generate-image.md) * [Generate video](/visualize/generate-video.md) * [Resolution and cost](/visualize/resolution-and-cost.md) * [First and last frame](/visualize/first-and-last-frame.md) * [How the visualizer thinks](/overview/concepts/how-the-visualizer-thinks.md) --- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question. Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter: ``` GET https://help.intangible.ai/visualize/ai-models.md?ask=&goal= ``` `ask` is the immediate question: it should be specific, self-contained, and written in natural language. `goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.