Every generative artist knows the feeling: the composition is perfect, the lighting is moody, the colors sing, and then you look at the hands. Six fingers. A wrist that bends the wrong way. A thumb that appears to be a separate creature. AI image generation has improved dramatically in almost every area, but hands remain the most famous failure mode in the industry. The good news is that this problem is not random, and it is not unsolvable. It comes from how diffusion models learn anatomy, and once you understand that, you can build a prompt strategy that dramatically reduces bad hands and a workflow to fix the ones that slip through.
Why AI Still Struggles With Hands
Diffusion models do not have a geometric model of the human body. They generate images by progressively removing noise from random pixels, guided by what their training data "statistically associates" with the words in your prompt. A hand is a small, highly articulated structure with enormous variation in pose, perspective, and occlusion. Training data contains billions of images of hands in every possible configuration, but the model learns an average, not a rule.
Fingers are particularly hard because they are small relative to the frame and they overlap constantly. When a model has to decide what five thin, similar shapes are doing at a given angle, small statistical errors compound into extra digits, fused fingers, or anatomically impossible joints. This is not a bug in your workflow; it is a structural limitation of the approach.
The second factor is tokenization and emphasis. In many models, the word "hand" competes with dozens of other words in your prompt for attention. If you mention hands only in passing, the model treats them as background detail and spends its limited capacity on the more heavily weighted elements of your scene. The fix, which this guide covers in detail, is to make hands a deliberate, described part of the image rather than an accidental one.
Start With Negative Prompts: What to Exclude
Negative prompts are the first line of defense, and most artists underuse them. In Leonardo AI, the negative prompt field acts as a filter: it tells the model which concepts to suppress during sampling.
Build a negative prompt library specifically for anatomy. Start with the basics: "bad anatomy, extra fingers, missing fingers, deformed hands, malformed hands, extra digits, fused fingers, disfigured hands, mutated hands, twisted fingers, claw hand, three fingers, four fingers, six fingers, seven fingers, extra arms, extra limbs, crossed eyes, bad proportions, deformed feet, missing limbs, amputee". These phrases map directly to the artifacts the model learned from flawed training images.
Write the negative prompt once, save it as a reusable preset, and apply it to every generation. It is far easier to prevent a bad hand than to fix one, and a solid negative prompt eliminates the most common artifacts before they appear.
One caveat: negative prompts are not magic. They reduce the probability of artifacts but cannot guarantee clean anatomy, especially in complex poses. Treat them as a necessary baseline, not a complete solution, and combine them with the positive techniques below.
Describe Anatomy in Positive, Specific Terms
While negative prompts remove the bad, positive prompts must actively construct the good. The model responds to explicit, weighted descriptions of hands when they are written as part of the scene.
Instead of letting hands exist implicitly, write them into the prompt with concrete language. Compare "a woman sitting at a cafe table" with "a woman sitting at a cafe table, her hands resting naturally on the tabletop, fingers gently wrapped around a coffee cup, relaxed open palm visible in the foreground". The second version forces the model to allocate attention to the hands and gives it a specific, low-ambiguity configuration to render.
Weighting helps. In Leonardo AI, you can emphasize key terms with parentheses and strength values. Boost phrases like "(detailed hands:1.3)", "(natural hand pose:1.2)", "(five fingers:1.2)" to push the model's attention toward clean anatomy. Do not overload the prompt with weights everywhere; a few strategic boosts on the anatomy terms work better than a prompt where everything is loud.
Precision matters in the pose description. "Hands" alone is vague; "left hand visible in profile, fingers extended slightly, thumb visible" is a pose the model can actually follow. The more you narrow the ambiguity, the fewer ways the model has to get it wrong.
The Action-Scene Technique
Descriptions of a static hand are useful, but the most common failures happen when hands are doing something. A hand holding a sword, reaching for a door, or gesturing in conversation all involve pose, perspective, and occlusion at once, which is exactly when models break down.
The action-scene technique solves this by embedding the hand in a clear narrative action with specific mechanics. Instead of "a knight holding a sword," try "a knight gripping a sword with both hands, right hand on the hilt, left hand supporting the blade, knuckles visible, fingers wrapped firmly around the leather grip". The action gives the model a coherent physical logic to follow, and the specifics reduce the number of anatomical decisions it has to invent.
If the action is complex, simplify the hand's job. A hand can be holding, resting, reaching, pointing, or gesturing. Pick one verb, describe the contact point, and describe the visible digits. If the model still struggles, change the framing: a close-up of a hand in a simple pose is far more reliable than a wide shot where hands are tiny and partially hidden.
Another practical trick is to move hands out of the worst conditions. Profile and three-quarter views are more reliable than straight-on palm shots, which compress the fingers into overlapping shapes. Backs of hands are simpler than palms. Occlusion is your friend: a hand partially behind an object, a sleeve, or a prop gives the model less anatomy to render, not more.
Use Reference Images and Style Control
Prompting alone will never match the reliability of a good reference. Leonardo AI supports image-to-image workflows and reference image features that give the model a concrete visual anchor.
For recurring characters, generate a reference sheet of the character's hands once, when the anatomy happens to come out clean, and reuse it in later generations. A character reference image that includes a good hand close-up trains subsequent generations toward the same anatomy. This is especially valuable for series, comics, and any project where the same character appears repeatedly.
For single images, use a style reference or an image prompt that provides a hand pose you like, and instruct the model to preserve the composition while changing the rest of the scene. This "pose borrowing" approach is more reliable than describing the pose from scratch.
Consistency features help too. When a platform offers a reference strength control, tune it carefully: too low and the reference barely matters, too high and the model copies unwanted details from the source. Start around a moderate setting and adjust based on how closely the result should match.
Iterate With Rerolls and Selective Editing
Even with the best prompts, some generations will come out wrong. The professional approach is not to restart from scratch but to iterate surgically.
Rerolling is the cheapest first step. Most platforms let you regenerate a new variation of the same prompt with one click, and the randomness of the process means an occasional clean result. Generate a batch, review them as a set, and keep the best anatomical candidates even if other aspects are imperfect, because fixing a hand is easier than rebuilding a composition.
Selective editing is the second step. Inpainting or region-based editing lets you mask the bad hand and regenerate only that area while keeping the rest of the image untouched. When you inpaint a hand, prompt the replacement region specifically: "natural human hand, five fingers, correct anatomy, matching skin tone and lighting". Keep the inpaint region tight; a loose mask invites the model to change the whole image.
Chain the fixes. Sometimes fixing one hand introduces a problem in the other, or the lighting in the inpainted region does not match. Fix, review, fix again, and stop when the result is good enough for the context. Perfect hands in a background detail matter far less than perfect hands in a close-up.
Post-Generation Fixes With External Tools
When the model absolutely will not cooperate, the last resort is a separate tool built for the task.
Hand-specific restoration models and face/body restoration add-ons are trained to repair anatomical errors. Running the image through one of these as a post-processing step can clean up minor digit problems without a full regeneration.
Classic image editors remain relevant. The clone stamp and healing brush are crude tools, but for a small isolated error, such as a single extra finger on a hand that is otherwise perfect, a two-minute manual fix beats an hour of prompt wrestling. If you are comfortable with more advanced tools, a simple transformation, skew, or liquify adjustment can correct a bent wrist or an odd joint angle.
For video projects, consistency requirements change the calculus. A hand error that appears for a single frame can be tolerated or patched differently than a hero shot that will be examined closely. Reserve the most expensive fixes for the frames where the hand is actually visible and important.
Ready-to-Use Prompt Templates
The fastest way to improve is to start from a working template. Here is a structure you can adapt:
Baseline positive prompt: "[subject and scene description], [hand action and pose described in detail], (detailed hands:1.3), (natural hand pose:1.2), (five fingers:1.2), [lighting], [camera and lens], [style]"
Baseline negative prompt: "bad anatomy, extra fingers, missing fingers, deformed hands, malformed hands, extra digits, fused fingers, disfigured hands, mutated hands, twisted fingers, claw hand, three fingers, four fingers, six fingers, seven fingers, extra arms, extra limbs, crossed eyes, bad proportions, deformed feet, missing limbs, amputee, blurry, low quality, jpeg artifacts"
Close-up hand prompt: "close-up of a human hand, palm facing camera, five fingers extended naturally, clean skin texture, soft studio lighting, macro lens, shallow depth of field, photorealistic, (detailed hands:1.4), (natural hand anatomy:1.3)"
Action prompt: "a hand gripping [object] with visible knuckles, fingers wrapped around [contact point], [specific digit positions], motion implied, cinematic lighting, (detailed hands:1.3), (correct hand anatomy:1.3)"
Swap the bracketed values for your scene, keep the anatomy terms boosted, and adjust the negative prompt as needed. Templates are starting points, not guarantees; the model still needs your eyes in the loop.
Frequently Asked Questions
Which models inside Leonardo AI are best for hands? It varies, but models with more photographic training data generally handle anatomy better than stylized ones. Test the same prompt across your available models, note which ones produce clean hands reliably, and default to those for character work.
Why does my negative prompt not eliminate all hand problems? Negative prompts reduce probability, they do not guarantee outcomes. Complex poses, occlusion, and tiny hands in wide shots will still fail sometimes. Combine negatives with explicit positive descriptions, references, and iteration.
Is it better to generate hands separately and composite them? Sometimes, yes. If a hand is the centerpiece of an image, generating it as its own crop with a dedicated hand prompt, then compositing it back, gives you far more control. It is more work, but it is the most reliable path for hero shots.
How many rerolls should I expect before getting clean hands? With a solid prompt and negative library, clean hands appear on the first or second batch more often than not for simple poses. Complex interactions may take several batches or selective editing. If a prompt consistently fails after many tries, change the pose or the framing rather than repeating the same attempt.
Does fixing hands get easier in video generation? Video inherits the image model's anatomy weaknesses, and errors can persist across frames, which makes them harder to fix. The mitigation is the same: consistent character references, simplified hand actions, and careful prompt design. Fix the anatomy problem at the image stage whenever possible.
Hands are the classic AI tell, but they are also the most teachable problem in generative art. The fix is a combination of prevention and repair: negative prompts to filter the obvious artifacts, specific positive descriptions to steer the anatomy, references to anchor the pose, and iteration with selective editing for the cases that slip through. Build those habits once, apply them to every character generation, and the six-fingered surprise will stop being the thing people remember about your work.
