Fixing Poorly Drawn Hands

Introduction Since the early days of AI art, specifically diffusion based art, AI has struggled the most with small details. Arguably, the most frustrating of these details is the human hand. I've heard more complaints about hand failures than most, if not all other failures combined. Not all models have these problems to the same extent. Some models, however are notorious for them. Generally the models got better with every generation. For example, SD (Stable Diffusion) was/is among the worst, characteristically drawing hands that looked like they went through a sausage grinder, and with ridiculous amounts of fingers. When SDXL came out, there were a lot of general improvements, but the difficulty with small details was not resolved, and if it was improved, it was not by much. Flux was a big leap forward, and comparing it to SDXL, Flux does a much better job with hands. That said, flux hands are still terrible. The Situation I made some images for the Purple Art challenge, and one of the pieces I really liked (below) had a terrible case of AI hands. It was surprising because I used GPT 1.5 to make these images, and it's generally very good with hands. This blog is about my search for the model that would fix those hands. I want to stress that several of these models are good at editing, especially the ones specifically designed for that task. Fingers has always been a special case. The Settings For every model I tried to keep all the parameters as much the same as possible. The broken image is 2:3 aspect ratio, and the prompt I used is Please fix the woman's fingers by ensuring there are only four fingers and one thumb on each hand . The Attempts Legend I am using emojis to quickly indicate how successful each model is. 🎉 — Complete success. 🤏 — The images shows signs of improvement. ❌ — There is no appreciable change. 🧩 — Wrong, but with a surprising result. 💢 — Frustratingly worse than the original. 🥀 — I've given up hope. 💩 — Utter failure beyond comprehension. GPT 1.5 ❌ ❌Attempt 01 GPT15 Low 2:3 — Negligible change. ❌Attempt 02 GPT15 Low 2:3 — Negligible change. Qwen / P-Image Edit 🤏⟹ ❌ 🤏Attempt 04 Qwen Image Edit 2:3 — Fixed, but hand moved. ❌Attempt 06 Qwen Image 2.0 2:3 — Negligible change. 🤏Attempt 07 Qwen Image 2.0 Pro 2:3 — Fixed, but way off style (sausage fingers). ❌Attempt 24 P-Image Edit 2:3 — Broken slightly worse. ❌Attempt 25 Qwen Image Edit Plus 1:3 — Fixed but dramatic hand repositioning. Seedream 🤏⟹ ❌ ❌Attempt 03 Seedream 5.0 Lite 2:3 — Negligible change. 🤏Attempt 26 Seedream 4 2:3 — Fixed, but moved, and fix is low quality ❌Attempt 27 Seedream 4.5 2:3 — Negligible change. Wan ❌ ❌Attempt 05 Wan 2.6 2:3 — Technically fixed, but hand posture failed (stiff). ❌Attempt 28 Wan 2.5 2:3 — Technically fixed, but both hands moved away and stiff Flux 2 🤏⟹ ❌ 🤏Attempt 11 Flux 2 Flex 2:3 — Better, but not quite fixed. ❌Attempt 13 Flux 2 Klein 2:3 — Negligible change. ❌Attempt 30 Flux 2 Dev 2:3 — Negligible change. ❌Attempt 31 Flux 2 Pro 2:3 — Negligible change. 🤏Attempt 32 Flux 2 Max 2:3 — A tiny bit better, but awkward hands. Flux 1 🤏⟹💩 🥀Attempt 12 Flux 1.1 Pro Ultra 2:3 — It's incredibly disappointing how poorly this worked. That said, none of the Flux 1 line (besides Kontext) were meant to do anything more than derive style information from reference images. 🤏 Attempt 14 Flux Kontext Pro 2:3 — Fixed, but hands moved. ❌Attempt 15 Flux Kontext Max 2:3 — Negligible change. 🧩 Attempt 29 Flux 1.1 Pro 2:3 — I'm quite impressed that the result was so close to the same image. Could have been a fluke. 💢 Attempt 33 Flux Dev 3:4 — Well, at least it drew the same basic elements. 💩 Attempt 34 Flux Krea Dev 3:4 — The emoji says it all. Individual Models 🤏⟹💢 🤏 Attempt 16 Grok Imagine Image 2:3 — Almost fixed. 🤏 Attempt 17 Hunyuan Image 3 Instruct 3:4 — Fingers look nice, but not well-connected to the hands. This one is close. 💢 Attempt 19 Minimax Image 01 9:16 — It drew a pretty girl's face and something that almost resembles fingers. 💢 Attempt 35 Gen4 Image 9:16 — It made a nice picture that barely resembles the original. Nano Banana 🎉⟹ ❌ ❌Attempt 20 Nano Banana 3:4 — — — Negligible change. 🎉Attempt 22 Nano Banana 2 2:3 — — Fixed. I would have preferred the hand slightly different, but this works. Almost as good as Nano Banana Pro. 🎉 Attempt 23 Nano Banana Pro 2:3 — Fixed. I would have preferred the hand slightly different, but this works. Ideogram v3 💢⟹💩 💩 Attempt 18 Ideogram v3 Balanced 9:16 — Umm... What do I even say about this? The primary set of fingers and thump are in the correct quantity. She has an extra set of thumbs? spikes? and those weird lumps, and wrist growths. 💢 Attempt 36 Ideogram v3 Turbo 9:16 — The resulting image is nice enough on its own. 💢 Attempt 37 Ideogram v3 Quality 9:16 — The resulting image is nice enough on its own. Conclusion The Winners There were two winning models: 1. Nano Banana Pro 2. Nano Banana 2 There were a few other models that came close, but unless you were comfortable with the composition changes, they were useless. Final Thoughts Obviously, several of these models are not designed to be used for editing, meaning the images are meant to be references only for style and concept. Of the ones that are, they each have their own limitations. Also, Hands are hard to fix for the same reason they are hard to draw in the first place. The fingers are very small.