Correction, 2026-10-01: this page said base Wan 3.0 reference-to-video, 34 credits for 5 seconds at 480p, exists for anchoring a shot to an approved still. On the canvas, base Wan 3.0 takes one start image (image-to-video, 24 credits for 5 seconds at 480p) and no reference images; its reference-to-video mode is available only through the 8frame MCP connector and starts from a reference clip, not a still. We also added the sources this page was missing. And we ranked models by how well they hold identity, from "Weak" to "Best available", and said Seedance 2.5 holds object identity longest, which we never measured; the ranking is replaced with what each model accepts as an anchor and a way to test morphing on your own shot.
Morphing is when an object in your clip gradually becomes something else: a hand merges into a cup, a chair leg joins a table, a face shifts identity partway through. It is the most characteristic failure of generated video and it has a specific cause that tells you exactly how to avoid it.
TL;DR
- Cause: the model has no persistent representation of objects. Its idea of what something is drifts across frames
- It scales with duration. Short clips give the drift less time to build
- It scales with object count. Fewer things in frame means less drift
- No model is immune, and we have not measured which morphs least. What differs, and can be checked, is the anchor each one takes: Seedance 2.5 up to 9 reference images, Veo 3.1 multi-reference 1 to 3, Kling v3 and Wan 3.0 one start image
- Test on your own shot: same start image, same prompt, 5 seconds, two or three models, compared frame by frame
- Some shots morph no matter what, and those are shots to cut rather than retry
Why it happens
A video model generates a sequence of frames that are locally coherent: each frame follows plausibly from the last. What it does not maintain is an internal model saying "this is a ceramic cup, it has these properties, it persists."
So across enough frames the model's implicit sense of what an object is drifts, and it renders something plausible-but-different. The drift is gradual, which is why it reads as morphing rather than as a jump.
This is also why the same shot can be clean at 5 seconds and broken at 15.
What reduces it
Shorter clips. The lever you control most directly. If a shot morphs at 10 seconds, generate it at 5 and cut. Two 5-second clips cost the same as one 10-second clip on most models and morph less.
Fewer objects. Each thing in frame is another thing that can drift. A single subject on a plain background gives the model the least to keep track of.
Less motion. Morphing accelerates with movement, because each frame differs more from the last.
A reference first frame. Image-to-video from a still you approved fixes the starting state, so the model begins from your design rather than its own guess. On Wan 3.0 that is a start image, from 24 credits for 5 seconds at 480p.
A stronger anchor. We have not measured which model morphs least, and it changes from shot to shot, so be wary of any ranking (this page used to have one). What does differ between models, and can be checked, is what you can hand the model to hold on to. Canon 8frame prices, a credit is $0.01 at pack rate:
| Model | Clip | Credits | Lengths on 8frame | What you can anchor it with |
|---|---|---|---|---|
| Wan 3.0 at 480p | 5s | 24 | 2 to 30 s | text, or one start image; no reference images on the canvas |
| Veo 3.1 Lite | 8s silent | 33 | 4, 6 or 8 s from text; 8 s with a start image | a start image, plus an end frame if you like; no reference images |
| Kling v3 Standard | 5s silent | 57 | 3 to 15 s | a start image, required; no reference images |
| Seedance 2.5 | 5s at 720p | 157 | 4 to 30 s | a start image and end frame, or up to 9 reference images (not both); one reference video, one audio track |
| Veo 3.1 Standard | 8s with audio | 448 | 4, 6 or 8 s from text; 8 s with a start image | a start image and end frame, or 1 to 3 reference images on the multi-reference option (8 s only) |
Read the last column as a choice of anchor, not a quality score. A start image fixes how the shot begins; reference images show the model the subject itself, which is the thing that drifts when a clip morphs. On the canvas, reference images go into Seedance (up to 9), Grok Imagine's reference mode (1 to 10, from 35 credits for 5 seconds at 480p), Happy Horse (1 to 9), Gemini Omni Flash (1 to 7), Kling O1 Reference (up to 7) and Veo 3.1 multi-reference (1 to 3). On Seedance, Gemini Omni Flash and Happy Horse, a start image and references do not combine: with both attached, the start image is dropped. Full limits are in reference image limits by model.
Static camera. Camera movement compounds morphing because the model is inventing parallax at the same time.
How to test morphing on your own shot
No ranking, ours included, tells you how your subject behaves. A short side-by-side does:
- Fix the inputs. One approved start image, one prompt, one aspect ratio. Change nothing but the model.
- Run it at 5 seconds on two or three models. From the same start image: Wan 3.0 at 480p (24), Kling v3 Standard silent (57) and Seedance 2.5 at 480p (70) is 151 credits, about $1.51, for three versions. Veo from a start image is fixed at 8 seconds; Veo 3.1 Lite silent is 33.
- Step through it frame by frame. Note the second where the object first changes. Look first where two objects touch, at hands, and at anything that passes behind something else.
- Extend the one that held. Re-run it at 10 seconds before you commit to a long take: holding at 5 does not mean holding at 10.
- If the subject needs references, test those the same way. The same reference images into each reference-capable model, same prompt, same length. A close-up face as a reference can trip Seedance's filter; see Seedance "flagged as sensitive" (E005).
The shots that morph regardless
Recognising these saves the most credits, because retrying them is how people burn a plan:
Two people interacting. Bodies pass through each other and both identities drift. The reliable failure.
Hands doing something specific. Extreme articulation plus constant self-occlusion. See which AI model handles hands best.
A crowd. Many faces, all drifting independently.
Camera moving through architecture. The model is inventing 3D it does not have, so straight lines bend and surfaces flex.
Long shots of a recurring character. Identity does not survive across a sequence without a locked reference, and even then it drifts at new angles.
Anything over about 10 seconds with a subject in it. The exception is abstract or textural material, which has nothing to keep consistent, so length costs it far less.
The long-clip exception
Worth knowing because it inverts the rule: Wan 3.0 generates up to 30 seconds in one pass, 142 credits at 480p on the base model. That length suits abstract and textural material, because there is no object whose identity can drift.
Smoke, water, foliage, light, drifting particles, gradients. A 30-second plate of any of those has no identity to lose. A 30-second shot of a person has a face and hands to lose.
The workflow that avoids it
- Block the shot at 5 seconds on Wan 3.0 at 480p, 24 credits. See whether it morphs at short length.
- If it morphs at 5 seconds, the composition is too complex. Simplify: fewer objects, less motion, static camera.
- If it holds at 5 and breaks at 10, generate at 5 and cut. Do not fight it.
- If the shot needs an object to persist through real motion, give the model more to hold: run the test above with reference images of the object on a reference-capable model (Seedance 2.5 takes up to 9, 157 credits for 5 seconds at 720p) and keep whichever version holds on your shot.
- If it is one of the shots on the list above, change the shot.
And since there is no video upscaler or repair step on this canvas, a morphing clip is a regenerate rather than a fix. The decision happens before you submit.
FAQ
Why does my AI video change objects halfway through? The model has no persistent object representation, so its sense of what something is drifts across frames.
Does a longer clip morph more? Usually: the longer the clip, the more frames the drift has to build over. Length is also the lever you control most directly.
Which model morphs least? We have not measured it, and it depends on the shot, so test it: the same start image and prompt at 5 seconds on two or three models, compared frame by frame. What you can compare before spending anything is the anchor each model takes: Seedance 2.5 takes up to 9 reference images, Veo 3.1 multi-reference 1 to 3, and Kling v3 and Wan 3.0 one start image and no reference images on the canvas.
Why is my 30-second abstract clip fine? Because there is no object identity to drift. Abstract material holds at length; subjects do not.
Sources
- 8frame prices: 8frame tool registry credit functions, read 2026-09-28; Wan 3.0 inputs on the canvas and Grok Imagine reference mode re-read 2026-10-01; Seedance 2.5 at 480p read 2026-10-01
- Inputs, reference image limits, start-image rules and lengths per model: 8frame tool configuration and per-model generation code, read 2026-10-01 (see which models work from text only and reference image limits by model)
Shorter clips, fewer objects, and cut the shots that cannot hold. The 8frame canvas is free and unlimited; generation is paid from $19/month.