Correction, 2026-10-02: this page ranked models on hands: Seedance 2.5 "strongest on physical behaviour", Veo 3.1 Standard "the quality ceiling", Kling v3 Standard "fails on complex motion", Imagen 4 Ultra "highest fidelity", Nano Banana Pro "stronger on stills". It also said how often hand shots work or fail ("performs far better", "fail reliably", "the hardest shot in generative video") and why models get hands wrong, as fact. We never measured any of this. The rankings are gone: the table now lists prices, lengths, resolutions and start-image requirements, the reasons come from a cited research paper, and the page explains how to test a hand shot cheaply. It also said video has no equivalent of editing a still; the practical route is to fix the hand in the start image before animating. Kling v3 Standard was listed without saying it needs a start image.
Correction, 2026-10-01: an earlier version gave Flux Edit a flat price of 7 credits and advised editing "with low strength". Flux Edit (FLUX.2 [pro] edit) bills by input plus output megapixels: about 7 for a 1 MP image, about 20 for a 2048 x 2048 image, around 36 for a 12 MP phone photo; Kontext Max is a flat 13. 8frame's image-edit nodes have no strength setting: the inputs are a prompt, the image and an aspect ratio.
Hands are a well-known failure point of generated images. Researchers describe the problem in diffusion models as "incorrect finger counts or irregular shapes", and in video the hand has to be right in every frame, not just in one.
We have not measured how often each model on 8frame gets hands right, so this page does not rank them. It gives you the facts per model, a cheap way to test your own shot, and the fixes you control.
TL;DR
- No model on 8frame guarantees correct hands, and we have not measured which does best, so test your shot before paying for a premium take
- Testing is cheap: a 5-second Wan 3.0 clip at 480p is 24 credits; Veo 3.1 Standard is 448 for 8 seconds with sound
- The fixes you control are compositional: frame hands out, give them something to hold, keep them still, or keep them small in the frame
- For stills, edit the hand rather than regenerate: Flux Edit is about 7 credits on a 1 MP image
- For video, fix the start image first: correct the hand in the still, then animate it
Why hands go wrong
The HandRefiner paper (Lu and others, 2023) puts it down to "the complex task of learning the physical structure and pose of hands from training images, which involves extensive deformations and occlusions." In practice that means:
Many possible poses. Hands bend into a huge range of shapes, and every finger has to end up somewhere anatomically possible.
Fingers hide fingers. Fingers overlap and cover each other, so the model has to infer what is behind what. A wrong guess shows up as an extra or a missing finger.
In video, every frame counts. A clip has to keep the hand right for its whole length, so check the full clip, not the first frame: a hand can change shape mid-motion.
That paper studied image models in 2023. The models on 8frame are newer, and we have not tested how much of this still applies to each of them.
The models, with facts instead of rankings
Canon 8frame prices, a credit is about $0.01 at pack rate:
| Model | What it makes | Credits | Start image |
|---|---|---|---|
| Wan 3.0 | video, 480p / 720p / 1080p, 2 to 30 s | 24 / 48 / 95 for 5 s | optional |
| Grok Imagine | video with sound, 480p / 720p, 1 to 15 s | 34 / 48 for 5 s | optional |
| Kling v3 Standard | video, 720p, 3 to 15 s | 86 for 5 s with sound, 57 silent | required |
| Seedance 2.5 | video, 480p / 720p, 4 to 30 s | 70 / 157 for 5 s | optional |
| Veo 3.1 Standard | video, 1080p, 4, 6 or 8 s | 448 for 8 s with sound | optional (8 s with an image) |
| Seedream 5 Lite | image | 6 | |
| Nano Banana Pro | image | 21 (42 at 4K) | |
| Flux Edit (FLUX.2 [pro] edit) | edit of an existing image | about 7 for a 1 MP image, about 20 at 2048 x 2048 | your image |
| Kontext Max | edit of an existing image | 13 | your image |
Price is not a quality ranking. The video prices here run from 24 to 448 credits, and we cannot tell you how much of that difference buys better hands. Your own test can.
Test your hand shot before the final take
- Make and check the start frame first. Generate the still (6 credits on Seedream 5 Lite, 21 on Nano Banana Pro), look at the hand, and fix it with Flux Edit if needed (about 7 credits at 1 MP). Then animate the corrected frame. The clip starts from a hand you have checked; the frames after it still need checking.
- Draft the motion cheaply. Run the shot on Wan 3.0 at 480p (24 credits for 5 seconds) or Grok Imagine at 480p (34) and look at the hand in every frame. A cheap draft tells you whether the shot is easy or hard; it does not predict what a different model will do.
- Draft on the model you will deliver on. Before its full-price take, run the prompt once at its lowest setting: Seedance 2.5 at 480p is 70 credits, against 157 at 720p.
- Compare on your own shot. Run the same prompt on two models and keep the better take. That is the only comparison that tells you about your shot.
The compositional fixes you control
Frame them out. Compose above the wrists, or crop so hands are outside the frame. Hands that are not in the shot cannot go wrong, and many shots do not need them.
Give them something to hold. A hand around a simple object with a clear shape (a cup, a phone, a box) shows fewer loose fingers, and the object covers some of them. Worth trying before an open, gesturing hand.
Keep them still. Less hand motion means fewer frames in which the hand can change shape. If the hand must be visible, let it rest.
Show part of the hand. A hand entering the frame, or partly behind something, leaves fewer fingers to get right.
Keep them small. At a distance a hand covers few pixels and errors are harder to see. A close-up of hands shows every flaw.
Plan for specific gestures. Counting on fingers, a specific sign, playing an instrument or typing all depend on exact finger positions. If your shot needs one, budget for several takes and check every frame, or film it.
For stills, edit rather than regenerate
If an image is right except for a hand, edit it rather than start over: Flux Edit (about 7 credits for a 1 MP image, about 20 at 2048 x 2048) or Kontext Max at a flat 13, with a targeted instruction that names the hand and what should change. The edit node takes a prompt, the image and an aspect ratio; there is no strength setting. An edit starts from your image, so check that nothing else changed. Regenerating, even at 6 credits, gives you a new image with a new composition, and the hand can come out wrong again.
Edit prices across tools are in AI image editing price by tool.
For video, fix the frame or cut around it
8frame has models that edit an existing clip (see edit existing video with AI), but we have not tested whether they repair a hand. The routes you control: fix the hand in the start image before animating (see the image-to-video workflow chain), regenerate, or cut around the frames where the hand goes wrong.
Where this matters
Product-in-hand shots. A person holding a product is a common social ad shot, and it is a hands shot.
A practical route: photograph the real hand holding the real product, then animate from that photo. The first frame then shows the real hand and the real product; the model generates every frame after it, so check those. Kling v3 requires a start image, and most other video models on 8frame accept one.
Craft and process content. Anything showing hands making something: film it.
FAQ
Which AI model does hands best? We have not measured it, so we do not rank models on hands. No model on 8frame guarantees correct hands. Test your shot on a cheap draft (Wan 3.0 at 480p is 24 credits for 5 seconds) and compare two models on the same prompt.
Are AI hands fixed in 2026? Not in a way we can promise. Check every image, and every frame of a clip.
How do I avoid the problem? Frame hands out, give them a simple object to hold, keep them still, or keep them small in the frame.
Can I fix a bad hand in a generated image? Yes, with Flux Edit (about 7 credits for a 1 MP image, more for larger ones) or Kontext Max (13) and a targeted instruction. For video, fix the start image before animating, regenerate, or cut around it.
Sources
- Lu and others, "HandRefiner: Refining Malformed Hands in Generated Images by Diffusion-based Conditional Inpainting", submitted 2023-11-29: "incorrect finger counts or irregular shapes" and "extensive deformations and occlusions": arxiv.org/abs/2311.17957, checked 2026-10-02
- Video prices: 8frame tool registry credit functions (calculateWan3Cost, calculateGrokImagineCost, calculateKlingV3Cost, calculateSeedance2Cost, calculateVeo31Cost), read 2026-10-02
- Image prices: Seedream 5 Lite 6 and Nano Banana Pro 21 (42 at 4K), 8frame tool registry, read 2026-10-02
- Flux Edit (FLUX.2 [pro] edit, billed on input plus output megapixels) and Kontext Max (flat 13): 8frame backend tool registry, FLUX_KONTEXT pricing branch and calculateFlux2Cost, read 2026-10-02
- Image-edit node inputs (prompt, image, aspect ratio; no strength setting): 8frame front tool registry, read 2026-10-01
- Lengths, resolutions and start-image requirements: 8frame front tool registry and video generation code, read 2026-10-02
Compose around hands, or photograph the real one.
Try it without signing up: the 8frame canvas opens on a real board. The canvas is free and unlimited; generation is paid from $19/month.