Skip to main content
MeiGen supports multiple AI models for image and video generation. Use the comparison table below to pick one, then scroll to its section for detailed specs.
Web-only models can differ from the public API list; see each model’s availability note. Availability and the current default come from List Models. Use a model only when it appears in the active list, and read is_default instead of assuming a permanent default. Prices below describe the listed configuration; confirm the displayed total before submitting.
Enhancing an existing video with Video Enhancement (Upscale) uses separate Skill pricing: 7 credits/sec at 1080P and 20 credits/sec at 4K. It is available on Web/App, rather than as a general video generation model in the public API/MCP model list.

​
Quick Comparison

✓ = available on the iOS/Android app and web. Web only = browser only for now. GPT image 1.5 was retired on 2026-04-22 (replaced by GPT Image 2.0). Midjourney V7 was replaced by V8.1 on 2026-05-09, then upgraded in place to V8.2 on 2026-09-02 (same model ID). Happyhorse 1.1 was retired on 2026-07-31. Seedream 5.0 Lite was retired on 2026-08-13 (Seedream 5.0 Pro and Seedream 4.5 remain available). Z Image Turbo was retired on 2026-08-21.

​
Image Models

​
GPT Image 2.5

For precise generation and editing. Select one public model ID and choose how to balance quality and speed. Credits per image: Sunburst and Flare cost the same at the same resolution and quality. Adding reference images or changing the aspect ratio does not add credits. Higher quality can take longer; there is no fixed completion-time guarantee. When selected, the configured starting settings are 2K Standard (6 credits) on desktop and 1K Standard (4 credits) on mobile, with Sunburst. Saved settings may differ. Both variants accept text-only prompts and reference images. For an edit, specify what to change and what to preserve. For a person, product, or style reference, explain the role of each image. Attaching a reference does not automatically switch variants. You can explicitly request a transparent background in your prompt. Whether the result has an alpha channel depends on intent detection and the generated output; use Cutout to remove the background from an existing image. GPT Image 2.5 uses welcome or purchased credits; daily free credits do not apply. 4K is paid-only. See API parameters for quality and modelVariant.

​
GPT Image 2.0

Credit matrix: All resolutions consume welcome or purchased credits (premium model — daily free credits don’t apply). 4K is paid-only. Aspect ratio does not affect pricing — credits depend only on resolution × quality. GPT Image 2 remains a separate model with its original three quality levels and prices. It supports instruction following, in-image text, and multi-element scenes. Its medium and high settings are not interchangeable with the same labels on 2.5. You can explicitly request a transparent background in your prompt. Whether the result has an alpha channel depends on intent detection and the generated output. Use Cutout when you need to remove the background from an existing image.
Stricter moderation: prompts involving real people, celebrities, brand logos, or violent/adult content may be rejected. Rephrase to drop specific names or try a different model.

​
Nanobanana 2.1

A separate model alongside Nanobanana 2, with a focus on sharper text rendering. Both remain available; use nanobanana-2.1 to select this version explicitly. Unlike Nanobanana 2’s flat 5-credit price, 2.1 costs 8 credits at 4K. Available on Web, iOS and Android. This is a premium model: daily free credits do not apply, and 4K is paid-only. API and MCP generation use purchased credits only. Call list_models for current availability and pricing.

​
Nanobanana 2

The most affordable Gemini model with the widest range of aspect ratios, including ultra-wide (8:1) and ultra-tall (1:8) options. Good for quick iterations and exploration. 4K is paid-only.
Content moderation: Nanobanana 2 has stricter content filtering than other models. Prompts that mention celebrities, brand names, real people, or face-swap scenarios are more likely to be rejected. If your prompt is flagged, try rephrasing to avoid specific names or use a different model.

​
Nanobanana Pro

The premium Gemini model with higher image quality. Better at complex scenes, fine details, and text rendering. Choose this when quality matters more than cost. 4K is paid-only.

​
Seedream 5.0 Pro

ByteDance’s flagship Seedream tier. Accepts up to 10 reference images, making it a strong pick for multi-subject composition and for keeping a character consistent across images. 2K is paid-only.

​
Seedream 4.5

The previous generation Seedream model. Still produces solid results, particularly good for product photography and marketing visuals. 4K is paid-only.

​
Grok Imagine Image 2.0

Credit matrix: xAI’s Grok Imagine 2.0 image model. Supports both text-to-image and image-to-image (up to 3 reference images for editing or guided generation), with a focus on precise editing, infographics, ads, and UI/UX mockups. A premium model — daily free credits don’t apply.

​
Midjourney V8.2

Midjourney’s photorealistic flagship. Returns 4 candidate images per generation (charged once). Choose 2K for native high-resolution output, High quality for finer detail at no extra cost. 2K is paid-only. Features:
  • Content and style reference modes
  • Advanced parameters (stylize, chaos, raw, etc.) for fine-tuning the output
  • Auto-translates non-English prompts to English
  • Native 2K output toggle
  • Quality (Standard / High) toggle

​
Flux 3 Image Klein

A basic text-to-image tier powered by Black Forest Labs through Vercel AI Gateway, fixed at 1K (about 1 MP). It costs 2 credits per image and remains eligible for daily free credits. Reference images and characters are not enabled in this MeiGen tier.

​
Agnes Image 2.5 Flash

The cheapest model on MeiGen at 1 credit per image, and the only basic-tier image model that accepts a reference image. Good for high-volume drafting and low-cost image-to-image iteration.

​
Video Models

​
HeyGen Video 1

Affordable video, optimized from MiniMax H3, with native audio.
Available on Web, API and MCP. Compatible App versions receive HeyGen through capability negotiation; older versions keep their existing model list. Reference-video availability follows the live model catalog and is available to existing HeyGen-compatible App versions.
First-frame and content-reference inputs use separate panels. Only the selected panel is submitted; switching panels preserves the other panel’s materials. Video or audio is submitted with content references. Video-only and image + video requests are supported. Reference video cannot currently be combined with separate reference audio; embedded video audio is allowed. This is a temporary MeiGen/Gateway compatibility restriction based on testing, not a published HeyGen model limit. Conflicting requests are rejected before charging; images + audio remain supported. HeyGen’s @character mentions are a MeiGen convenience: saved character images are submitted as ordinary content references, without Seedance assets. Reference up to 3 characters; one character uses up to 2 images, while multiple characters use 1 image each. Character images and uploads share the 9-image and 12-item limits after deduplication and use content-reference pricing. Character mentions select the content panel; the first-frame draft is preserved but not submitted.

​
Pricing

Credits = (output seconds + the sum of each reference video’s duration rounded up to whole seconds) × the rate for the submitted input mode. With no reference video, only output seconds count. The interface selects the rate automatically. Audio does not add billable seconds, but requests with content images, videos or audio use the higher content-reference rate. There is no minimum-seconds multiplier. For example, 5 seconds at 768p costs 15 credits with text or a first frame, and 30 credits with content references. A 4.1s reference video + 5s output at 768p costs (5 + 5) × 6 = 60 credits. This is a premium model and uses purchased credits only.

​
Image and prompt handling

Images reuse MeiGen’s shared upload component: JPEG/PNG/WebP/GIF, an 8000px-per-side upload guard, compression targeting 2MB and 2048px when needed, and a 10MB upload-gateway cap. AVIF is converted before upload. These are MeiGen upload settings, not published HeyGen pixel limits. Library/gallery images are also checked before charging: readable image data, a verifiable size no larger than 16MB, and a direct HTTPS URL without redirects. HeyGen accepts up to 32,000 prompt characters. MeiGen reserves space for protected reference instructions and native reference labels before applying the shared long-prompt compression. Within the remaining budget no compression is needed; beyond it the compressor targets 50 characters below that budget. On compression failure the body is shortened to fit, while reference instructions are protected and checked. The model receives labels such as <Picture 1> and <Audio 1>; use the editor’s @ mentions to insert them. Sources: HeyGen model guide and AI Gateway model. MeiGen’s current UI limits and submit-time credit quote are authoritative.

​
Seedance 2.5

Seedance 2.5 is the long-duration Seedance model on MeiGen. It supports text-to-video, first/last-frame image-to-video, up to 30 content-reference images, and reference-video continuation. Content images, reference videos and reference audio can all be used in the same request and are three independent budgets. Frame references and reference-video continuation use adaptive ratio so the output follows the input media; text-to-video and content-reference requests can use an explicit supported ratio. Editing versus continuation (Seedance 2.5). To continue a clip, describe what happens next. To edit the existing clip (for example, remove lettering or replace a person), use a 4–30 second source and set duration to that source’s rounded whole-second length. If multiple videos could be selected for editing, every source must match the requested duration; using one source avoids ambiguity. The backend may recover an explicit edit-duration rejection once by inheriting the source duration, only when this preserves the already billed output seconds. A mismatch fails and follows normal refund settlement; it never authorizes extra charging. Keep passing a positive integer: duration=-1 and omni_reference_task_type are internal provider parameters, not public API inputs. After a failed/refunded job, correcting the inputs requires a separately authorized new request ID. Retrying an uncertain submission keeps its original ID and exact inputs.

​
Pricing

Direct generation bills only the selected output duration. With one or more reference videos, billing uses the sum of their durations: billable seconds = max(sum of reference video durations + selected output duration, ceil(selected output duration × 5 / 3)) The minimum-billable floor is 7 seconds for a 4-second output, 9 for 5 seconds, and 50 for 30 seconds. Output length is always the selected 4–30 seconds; the billing floor does not lengthen the generated clip. Reference audio is free — its duration never enters the sum, and a request with audio but no reference video is billed at the direct rate.
If generation fails, the full credit amount is automatically refunded.

​
Seedance 2.0

Seedance 2.0 is the three-tier Seedance option. It supports text-to-video, image-to-video (first/last frame or content reference), and video continuation. All outputs include AI-generated audio. The adaptive aspect ratio automatically matches the reference image/video dimensions. Both Seedance 2.0 and 2.5 support character assets: save a character once, then reference it with @ in your prompt to keep the same subject across shots. Two reference-image modes: first/last frame (default, up to 2 images) sets the opening and/or closing frame; content reference guides style or subject instead of fixing a frame position and accepts up to 9 images. Content images, reference videos and reference audio can all be used in the same request and are three independent budgets. Three tiers, picked from the tier switch (Mini / Fast / Pro) in the video sidebar. Same Seedance 2.0 model, same params, just different rendering quality and price:
  • Mini (default): cheapest and most lightweight; 480p / 720p
  • Fast: quick turnaround at 480p / 720p
  • Pro: higher fidelity, with native 1080p and 4K rendering (4K is direct-generation only)
All three tiers support multi-clip reference-video continuation and reference audio — continuation output tops out at 1080p even on Pro.

​
Pricing

Per-second. Rate depends on tier (Mini / Fast / Pro) × whether a reference video is uploaded: Mini and Fast only support 480p / 720p. Pro adds native 1080p and 4K for direct generation; with a reference video, Pro is limited to 720p / 1080p (no 480p rate). With-reference-video billing rule (all tiers): billable seconds = max(sum of reference video durations + your selected duration, min billable) Output length is always your selected duration (4–15), independent of the reference videos’ lengths. Reference audio is free and never enters the sum; attaching audio without any reference video keeps the direct rate.

​
Examples

If generation fails, the full credit amount is automatically refunded — including all per-second charges.

​
Grok Video 1.5

MeiGen offers Lite and Standard tiers through Vercel AI Gateway. Both require a starting (first-frame) image — text-only generation is not supported in this integration. Output follows the image’s aspect ratio and includes generated audio. Generation time varies by tier, resolution and provider load.

​
Pricing

Total credits = the selected tier’s rate at the selected resolution × output seconds. Default settings are Lite, 480p, 4 seconds — 16 credits. Lite remains a premium model; daily free credits do not apply. Examples: API and MCP keep the model ID grok-video. Set tier to lite or standard; omitting it uses Lite. Set resolution to 480p, 720p, or 1080p. Read current defaults and pricing from List Models; do not send the upstream provider’s model ID as modelId.
If generation fails, the full credit amount is automatically refunded.

​
Veo 3.1

Google’s video generation model. Creates short clips with AI-generated audio synced to the prompt. Auto aspect ratio infers the output ratio from your prompt or reference image. Pick this for premium visual quality with native audio; for longer durations (>8s), use Seedance 2.5. Two tiers — pick one from the tier switch (Fast / Pro) in the video sidebar. Same Veo 3.1 model, same params, just different rendering quality and price:
  • Fast (default) — quick turnaround, balanced quality
  • Pro — higher fidelity, longer rendering time

​
Pricing

Per generation. Price depends on tier (Fast / Pro) × duration (4 / 6 / 8s). All resolutions (720p / 1080p / 4k) share the same price — picking a higher resolution costs no extra credits. Two caveats: 4k requires a paid account, and it takes noticeably longer to render.

​
Examples

If generation fails, the full credit amount is automatically refunded.

​
Agnes Video 2.5 Flash

Temporarily unavailable since 2026-09-22. This model is currently disabled: it is not returned by List Models and generation requests for agnes-video-2.0 are rejected with 400. The details below describe its behaviour when enabled and are kept for existing integrations. Check List Models for current availability.
The only basic-tier video model — and the only one your daily free credits cover — while it is enabled. Supports text-to-video and single-image first-frame image-to-video at a fixed 5 seconds / 720p. For longer clips or higher resolutions, use Seedance 2.5 (up to 30s and 1080p); for 4K, use Seedance 2.0 Pro or Veo 3.1.
Image-to-video ignores the requested aspect ratio — the output follows the first frame. Images you generated or uploaded on MeiGen are cropped to your selected ratio at the CDN edge first, so the result matches your choice. Images from anywhere else (API callers may pass any public URL or data: base64) are used as-is, and the output ratio follows that image.

​
Pricing

Duration is fixed at 5 seconds, so every generation costs 10 credits.
If generation fails, the full credit amount is automatically refunded.

​
Known Limitations

This model is more sensitive to content involving real people, celebrities, brand logos, and face-swap requests. Prompts that work on other models may be rejected here. If you get a content violation error, rephrase your prompt to remove specific names or try a different model.
This base model generates images from text only. The reference image upload section is hidden while it is selected. If you need reference images, switch to another model.
Midjourney V8.2 accepts only 1 reference image, while other image models accept between 3 and 16. In exchange it offers more control over how the reference is used (content vs. style reference).
Unlike other models that return a single image, Midjourney V8.2 returns 4 candidate images per generation. All candidates are saved and can be browsed in the image detail dialog. You are only charged once (12 credits at 1K / 18 credits at 2K) for all 4 images.
GPT Image 2.0 applies stricter moderation than other models to real people, celebrities, brand logos, and adult/violent content. Rephrase your prompt to drop specific names, or try a different model.
Seedance bills per second, not per image — and uploading a reference video switches to a separate rate table with a minimum-billable floor. With several reference videos, the billed seconds use the sum of their durations plus your selected output duration, so each extra clip raises the total. Reference audio is free and never enters that sum, and a request with audio but no reference video stays on the direct rate. See the Seedance 2.0 section above for the full rates and formula.
On Seedance 2.0, reference audio only works alongside at least one reference image (a saved character counts) or one reference video — audio on its own is rejected before any credits are deducted. Seedance 2.5 has no such requirement and accepts audio-only requests.
Seedance 2.5 has one quality level and supports 480p / 720p / 1080p. Do not send tier; use Seedance 2.0 Pro when you specifically need 4K.
The Mini tier (the default) is the lightweight option, capped at 720p. For 1080p or 4K output, switch to Pro — note that reference-video continuation on Pro tops out at 1080p (4K needs direct generation with no reference video).
Grok Video 1.5 is image-to-video only — you must upload a starting (first-frame) image. Pure text-to-video is not supported. The output aspect ratio follows your input image. If you want text-to-video, use Seedance 2.5, Seedance 2.0, or Veo 3.1.
Both tiers support 480p, 720p and 1080p. Lite is the default; price depends on both tier and resolution. For 4K, use Seedance 2.0 Pro or Veo 3.1.
Veo 3.1 only supports 16:9 / 9:16 (plus Auto). Square (1:1), 4:3 / 3:4 portrait, and other ratios are not available. For wider ratio support, use Seedance 2.5 or Seedance 2.0.
720p / 1080p / 4k all cost the same on Veo 3.1, so 4k costs no extra credits. Two things to know: 4k is locked for free accounts (selecting it opens the upgrade dialog), and rendering takes longer — 4k typically takes 6–8 minutes versus 1–4 minutes for 720p / 1080p. The same 4k lock applies to Seedance 2.0 Pro.