What is Reference image? Definition, key points and related tools
An image passed with the prompt to lock a character's look, a scene's space or a prop's material. Limits vary (MiniMax H3 up to 9, Wan 3.0 up to 10, Veo 3.1 up to 3, Qwen-Image 3.0 1–3, Seedream 5.0 Pro up to 10 and 5.0 Lite up to 14); MiniMax H3 and Wan 3.0 cannot combine a first frame with references in one request.
What a reference image transfers
A reference is conditioning, not an instruction. The model cannot tell which properties of the picture you meant, so everything visible is a candidate to copy: face, pose, light, expression, background, printed text. Vendors split references by job. Wan 3.0 names character or object references, style references and camera-movement references taken from a video; Kling's Element Library covers characters, animals, props, costumes and accessories, scenes and effects; Google's Veo 3.1 guide lists scene, character, object or style.
In AI short drama three jobs matter: identity (a character sheet plus a face crop), place (a scene plate) and key props. The reference says what they look like; the prompt says who does what.
Input specs that decide whether a reference is accepted
Official limits for image inputs, checked 2026-09-12:
| Model | Formats | File size | Pixels per side | Aspect ratio (w/h) |
|---|---|---|---|---|
| MiniMax H3 | JPG, PNG, WEBP, HEIC, HEIF | up to 30 MB | 256–5760 | 0.4–2.5 |
| Seedance 2.x (BytePlus) | JPEG, PNG, WEBP, BMP, TIFF, GIF, plus HEIC and HEIF on 1.5 pro and later | under 30 MB | 300–6000 | 0.4–2.5 |
| Wan 3.0 | JPEG, PNG without transparency, BMP, WEBP | up to 20 MB | 240–8000 | up to 8:1 |
| Vidu reference-to-video | PNG, JPEG, WebP | 50 MB per image, 20 MB per request | from 128 | within 1:4 to 4:1 |
In practice: a 9:16 sheet (0.56) passes everywhere, while a tall turnaround strip at 1:3 (0.33) is rejected by H3 and Seedance, so supply extra angles as separate images. Wan 3.0 does not take transparent PNGs, so flatten cut-outs onto white before upload. H3 and Seedance cap the whole request body at 64 MB and advise URLs rather than Base64 for large files.
Runway's guidance for character references sets quality floors on top: at least 1,024 pixels on the shortest side, sharp focus on the face, even lighting, one person, a front or three-quarter view and a simple background.
Binding, order and cost
References are addressed by position or by name: Wan 3.0 prompts refer to Image 1 and Image 2 in the order of the media array, Vidu uses @subjectname, Runway uses @tags and Kling binds elements. State the binding in the text (Image 1 is Mara, full body; Image 2 is Mara's face) and keep the same order in every segment of a scene, so the storyboard and the array never disagree.
- MiniMax H3: up to 9 images; the first 5 are free, then $0.04 each. At 768P ($0.08 per second) a 10-second segment costs $0.80, and a full set of nine adds $0.16, or 20%.
- Veo 3.1: up to 3 images. Attaching any fixes the clip at 8 seconds, and with references the person setting is limited to adults, which rules out referenced child characters on that route.
- Seedance: 1–9 images on 2.0, 1–30 on 2.5.
Fill slots in the order the viewer checks: face crops of everyone in frame, then the plate, then props.
Common mistakes and their causes
- Poster references. Mood lighting, colour grade and a dramatic pose are copied into every shot, where they fight the storyboard's own lighting.
- Text and watermarks inside the image. They are pixels like any other and can reappear. Seedream's API adds a watermark unless watermark is set to false, so switch it off when generating reference art.
- Temporary links. Seedream output URLs expire after 24 hours and Veo keeps generated videos for 2 days. Store your own copies before episode two needs them.
- Real-person photos. Seedance 2.x refuses directly uploaded references containing real human faces; likeness rights apply regardless of what a model accepts.
- Mixing modes. On H3, Seedance and Wan, a first or last frame cannot be combined with references in the same call, so a frame-pinned segment has no reference at all (see first / last frame).
Key figures and sources
| Figure | Value | Source | Date |
|---|---|---|---|
| MiniMax H3 image input | JPG/PNG/WEBP/HEIC/HEIF, up to 30 MB, 256–5760 px, aspect 0.4–2.5; up to 9 references | platform.minimax.io | 2026-09-12 |
| Seedance 2.x image input (BytePlus ModelArk) | under 30 MB, 300–6000 px, aspect 0.4–2.5; 2.0: 1–9 references, 2.5: 1–30 | docs.byteplus.com | 2026-09-12 |
| Wan 3.0 image input | up to 20 MB, 240–8000 px, aspect up to 8:1, no transparency; up to 10 references | www.alibabacloud.com | 2026-09-12 |
| MiniMax H3 reference pricing | first 5 images free, then $0.04 each; video $0.08/s at 768P | platform.minimax.io | 2026-09-12 |
| Runway character-reference guidance | at least 1,024 px on the shortest side; one person; even lighting; front or 3/4 view | runway.com | 2026-09-12 |
FAQ
Can one reference image hold two characters?
Avoid it. Runway's guidance asks for a single character per reference, and position-based binding (Image 1, Image 2) cannot say which face in a shared image belongs to whom. Give each character their own images and bind them by name in the prompt.
Why was my reference rejected?
Check it against the target model's window: file size, pixels per side, aspect ratio (0.4 to 2.5 on H3 and Seedance), transparency on Wan 3.0, and real human faces on Seedance 2.x. Large Base64 payloads can also exceed the 64 MB request cap on H3 and Seedance; send public URLs instead.
Do more references always mean more consistency?
No. Each image competes for the model's attention, caps are low on some models (Veo 3.1 takes three), and on MiniMax H3 every image beyond five costs $0.04. Cover the faces of everyone in frame first, then the location, then props; see character consistency for per-segment budgets.
Sources
- https://platform.minimax.io/docs/api-reference/video-generation-v2-create
- https://platform.minimax.io/docs/guides/pricing-paygo
- https://docs.byteplus.com/en/docs/ModelArk/1520757
- https://docs.byteplus.com/en/docs/ModelArk/1541523
- https://www.alibabacloud.com/help/en/model-studio/wan3-video-generation-api-reference
- https://www.alibabacloud.com/help/en/model-studio/wan3-video-generation-guide
- https://platform.vidu.com/docs/reference-to-video
- https://kling.ai/quickstart/klingai-element-library-3-user-guide
- https://ai.google.dev/gemini-api/docs/veo
- https://cloud.google.com/blog/products/ai-machine-learning/ultimate-prompting-guide-for-veo-3-1
- https://runway.com/resources/ai-character-references-tips