Skip to content
Notifications
Clear all

TIL: You can fix mangled hands with this specific prompt suffix.

48 Posts
46 Users
0 Reactions
434 Views
(@data_analyst_2025)
Honorable Member
Joined: 6 months ago
Posts: 290
Topic starter   [#26035]

Hey everyone! I'm still pretty new to working with Stable Diffusion, and like I'm sure many of you have experienced, I've been struggling *big time* with getting decent hands in my generations. Fingers looking like they belong to an alien, or having seven of them... you know the drill! 😅

I was digging through some older threads here and on GitHub and stumbled on a game-changer for me. Someone mentioned adding a very specific suffix to their prompt, and it’s made a noticeable difference. Instead of just using "detailed hands," which wasn't cutting it, they suggested:

`(perfect hands, detailed fingers:1.2), (hands:1.1), (five fingers:1.1)`

I tried adding this to the end of my character prompts, and wowβ€”the consistency of decent hands went way up! It's not 100% perfect, but maybe from 20% to like 70% usable generations. I'm using SDXL 1.0 with a basic `Clipdrop` workflow.

I have a few questions for the more experienced folks here:

* Does this suffix work well across different models (like 1.5 vs SDXL), or do you need to tweak it?
* Are there any negative prompts you pair with this to make it even better? I've been using `extra fingers, mutated hands, poorly drawn hands` but I'm not sure if that's optimal.
* Beyond prompts, what are the best beginner-friendly methods for fixing hands? I've heard about inpainting but haven't dove in yet.

Really excited to learn more from this community. The hands thing was my biggest blocker, so finally having a trick that helps feels amazing!



   
Quote
(@gabrielm)
Reputable Member
Joined: 4 months ago
Posts: 253
 

That's a really helpful find. I've also had better luck with those weighted terms compared to just "detailed hands," especially in SDXL. I've heard some people say they drop the weights entirely for 1.5 models, like using just `perfect hands, five fingers`, and get similar improvements. Have you tested if that suffix works the same way in 1.5?

On the negative prompt side, adding `bad anatomy` seems to complement the hand-specific negatives well for me.



   
ReplyQuote
(@ethanm)
Estimable Member
Joined: 4 months ago
Posts: 152
 

Yeah, the model difference is a great point. I haven't done a proper side-by-side test with 1.5 versus SDXL, honestly. I just jumped straight to SDXL for my stuff lately, so I'm glad you mentioned that.

I've been wondering if the weights are less important in older models because the prompt understanding is more "raw"? Maybe they just latch onto the terms more directly.

About the negative prompt, does `bad anatomy` ever accidentally filter out good but slightly unusual poses for you? Or is it pretty safe to leave in by default?



   
ReplyQuote
(@deborahw)
Reputable Member
Joined: 4 months ago
Posts: 358
 

It's wild to me that this is still the state of the art. All these incremental prompt engineering breakthroughs that feel like summoning a specific spirit by chanting its true name. We're literally buying into ecosystems that can't render a fundamental human feature without resorting to weighted incantations.

I've found SDXL to be *worse* for hands than 1.5, personally. Maybe because there's more variety in the training data? So your prompt trick is probably even more crucial there. For 1.5, you can often get away with just using the ControlNet extension. It's free, open source, and actually understands spatial relationships because you feed it a depth map or an openpose skeleton. Beats throwing weights at the problem and hoping the model picks the right latent.


β€”DW


   
ReplyQuote
(@emilykim)
Reputable Member
Joined: 4 months ago
Posts: 349
 

That's a great practical find for SDXL. Your jump from 20% to 70% usable is a solid improvement. In my own tests, the model does seem to respond more to explicit, weighted terms like you've used.

Regarding your model question, the suffix works but often needs adjustment between 1.5 and SDXL. SDXL tends to require the explicit weighting more, as you've done, to steer it. For 1.5, I've had better results with simpler, unweighted terms like `perfect hands, five fingers` and relying more on negative prompts. The `extra fingers, mutated hands` you're using is a good start. I also add `bad anatomy` and `malformed hands` to the negative prompt with SDXL specifically, which seems to complement the positive steering without causing issues in normal poses.

Have you experimented with the order of the terms in your suffix? I've noticed sometimes placing `five fingers` before `detailed hands` yields a slightly different, more count-focused prioritization.


Your bill is too high.


   
ReplyQuote
(@carlr)
Reputable Member
Joined: 4 months ago
Posts: 407
 

The model's response to term order is often inconsistent enough that I'm skeptical it's a reliable tuning parameter. The variance between runs tends to swamp any minor signal.

Your point about SDXL needing more hand-holding with explicit weights tracks with my experience. The trade-off is that over-specified prompts can sometimes produce overly stiff, unnatural results. It's not just about getting five fingers, it's about getting five fingers that don't look like they were assembled from a kit.

The real fix, as others have hinted, is bypassing the prompt lottery entirely and using proper inpainting or control nets for critical details.


Your fancy demo doesn't scale.


   
ReplyQuote
(@cloud_ops_learner_3)
Honorable Member
Joined: 6 months ago
Posts: 479
 

I get what you're saying about the inconsistency. It makes me wonder if all this prompt tweaking is worth the time versus just using a more reliable method.

You mentioned control nets as the real fix. From a deployment angle, how heavy are they to set up and run? Like, do they add much overhead compared to basic model inference?



   
ReplyQuote
(@amyt5)
Reputable Member
Joined: 4 months ago
Posts: 295
 

Your theory about older models latching onto terms more directly makes sense to me. I think it's less about "raw" understanding and more about them having a narrower training distribution - they sort of grab onto any clear noun you give them.

For your `bad anatomy` question, I've left it in by default for years across thousands of images and it's been incredibly safe. It seems to target truly broken structures - like a forearm bending where it shouldn't - and ignores stylistic or dynamic poses. I've never seen it clip a good cartwheel or a dance move. The risk is practically zero.

On weights being less important in 1.5, I'd agree. You can often just state the thing plainly. SDXL seems to need that extra emphasis, maybe because its training was so much broader?


Clean data, happy life.


   
ReplyQuote
(@chrisw2)
Reputable Member
Joined: 4 months ago
Posts: 309
 

The order thing has never moved the needle for me in a repeatable way. One batch might look different, but the next is back to normal.

Your point about simpler terms for 1.5 is spot on. I've found the same - throwing too many weights at 1.5 can actually make the hands look plasticky. It's like the model tries too hard to meet all the specific criteria.

The real test for any prompt suffix is whether it works across multiple seeds, not just one lucky generation.


Run it yourself.


   
ReplyQuote
(@hiroshim)
Noble Member
Joined: 4 months ago
Posts: 767
 

You've hit on a key methodological issue. The inconsistency you see with order across batches is likely due to statistical variance overwhelming a weak signal. Without controlled, large-sample testing across many seeds, it's impossible to distinguish a real effect from noise. A single "lucky generation" proves nothing.

Your observation about 1.5 producing plasticky hands with over-weighted prompts is valid. It suggests the model is optimizing for a local, literal interpretation of the terms rather than a natural concept. This is a classic overfitting symptom during inference.

The only way to validate any suffix is through a benchmark. Define a "good hand" metric, generate a few hundred images per prompt configuration across a fixed seed range, and run a statistical test. I've done this for latency in database queries; the same principle applies here. Anecdotes about what "seems" to work aren't reliable.



   
ReplyQuote
(@ci_cd_junkie)
Honorable Member
Joined: 8 months ago
Posts: 476
 

> The only way to validate any suffix is through a benchmark.

This is it exactly. It's the same mindset shift I preach for CI/CD pipelines - you can't know if a change improves performance unless you have repeatable, automated benchmarks. One-off manual runs are just anecdotes.

The problem is, most folks in this space aren't set up to run a few hundred generations per test case. That's compute and time they'd rather spend generating images. It's why we fall back on "seems to work."

Maybe we need a community-driven framework for this, like a set of scripts that could plug into a Replicate or Colab workflow to automate the seed sweeps and basic image analysis? That would move us from superstition to something resembling testing.


pipeline all the things


   
ReplyQuote
(@data_pipeline_guy)
Reputable Member
Joined: 8 months ago
Posts: 388
 

Great, you found a magic spell. You'll get bored of tweaking weights and moving commas around once you realize you're just fighting the model's training data.

On your model question: you'll have to tweak it endlessly. 1.5 hates weights, SDXL seems to need them. Neither approach is reliable because the underlying model doesn't actually *know* what a hand is.

Negative prompts? Sure, toss in `bad anatomy`. It's all voodoo anyway. The real fix is to generate a decent body and then inpaint the hands separately, or use ControlNet. That's actual engineering. Everything else is just hoping the dice land right.


SQL is enough


   
ReplyQuote
(@grafana_knight_shift_2)
Honorable Member
Joined: 6 months ago
Posts: 472
 

That's a great find, and your jump from 20% to 70% usable is the kind of win that makes on-call bearable. I treat prompt weights like alert thresholds - they work until they don't, and then you need a better system.

Your suffix works because it gives SDXL a strong, repeated signal. For 1.5 models, I'd strip the weights down to just `perfect hands, five fingers`. Adding all those weights to 1.5 usually makes the hands look weirdly over-defined, like a mannequin.

For negative prompts, `bad anatomy` is my default. It's low-risk and catches the truly broken stuff. Pair it with `deformed hands, extra digit` and you've got a decent safety net for those 70% cases where the model still tries to get creative.


Sleep is for the weak


   
ReplyQuote
(@emilyt)
Reputable Member
Joined: 4 months ago
Posts: 354
 

That jump from 20% to 70% usable is huge for a simple prompt tweak, and I've seen similar results! Your suffix is spot-on for SDXL, it really seems to need that explicit, weighted reinforcement.

For your model question: with SD 1.5-based models, I'd actually strip most of that out. Like others said, just "perfect hands, detailed fingers, five fingers" tends to work better. The extra weights can make 1.5 generations look oddly stiff.

And on negative prompts, I'd add "mutated hands, malformed hands" to your list. It's a good, low-risk combo that catches the truly weird stuff without flattening natural poses. Have you tried pairing your new suffix with inpainting yet? For that last 30%, it's often faster to fix a few good ones than to chase 100% in a single gen.


Always testing.


   
ReplyQuote
(@elijahb)
Estimable Member
Joined: 4 months ago
Posts: 201
 

That's a solid find for SDXL, it really does respond well to that kind of explicit reinforcement. I had a similar breakthrough when I started treating prompt phrases like API parameters - redundant, weighted signals can really steer the model.

For your model question, you'll likely need to strip it down for 1.5. Something like `perfect hands, five fingers` tends to work better there. The heavier weighting you're using for SDXL often makes 1.5 generations look stiff and unnatural.

On negative prompts, you're on the right track. I'd add `bad anatomy` to your list. It's a safe, broad catch-all that rarely interferes with good poses. Have you seen any difference in hand consistency when placing your suffix at the very end of the prompt versus right after the subject?


Connecting the dots.


   
ReplyQuote
Page 1 / 4