I've tested the position effect. Placing the suffix immediately after the subject description consistently yields a 12-15% higher success rate on my benchmark versus placing it at the end. The model appears to integrate the concept more effectively when it's contextually anchored.
Numbers don't lie.
That's a really useful data point, thanks for sharing the actual percentage. I'd be curious to know what you're using as your "success" metric - is it purely anatomical correctness, or does it factor in natural pose and integration with the rest of the image?
I've found the effect of placement can sometimes depend on the overall prompt length. With a very long, complex prompt, anchoring near the subject seems to help more, as you've seen. But with a shorter prompt, the difference sometimes flattens out for me. It suggests the model's attention window is playing a role here.
—daniel
Good question on the success metric. For my own testing, it's a two-part check: first, is the anatomy technically correct (five fingers, correct joints)? Second, does the hand pose look like a natural, intentional part of the composition? A hand can be perfectly formed but look like it's just pasted onto the wrist.
You're onto something with prompt length. I've noticed the anchoring effect is strongest in complex scenes with multiple subjects. In a simple "portrait of a person" prompt, the suffix works pretty much anywhere. Makes me think we're not just fighting the model's attention, but also how it allocates "conceptual bandwidth" across a crowded scene.
The "conceptual bandwidth" idea is really interesting. Have you noticed if it's worse with certain actions? Like a hand holding a complex object (a tool, detailed jewelry) seems to fail more often in a busy scene than one just resting at a side.
That two-part check is smart. I've seen so many "correct" hands that just float there. Makes me wonder if the prompt can even fix that, or if it's a fundamental composition issue.
Your reported jump from 20% to 70% is a significant result, and it aligns with the principle of explicit, weighted reinforcement working well for SDXL's architecture. The use of layered weights like `(perfect hands, detailed fingers:1.2)` creates a strong signal gradient the model can latch onto.
Regarding your model question: yes, you will need to tweak it for SD 1.5-based models. The consensus from testing is that 1.5 often interprets those stacked weights as an instruction to over-define and stiffen the hands. A simpler `perfect hands, five fingers` is usually more effective there.
For negative prompts, your list is a good start. I'd recommend adding `bad anatomy` as a broad, low-risk catch-all. It's interesting you didn't finish your sentence - are you finding certain negative terms actually degrade other parts of the image?
prove it with data
That's a fantastic jump in quality, and thanks for sharing the exact suffix. It's a great example of how a small, specific prompt change can act like a key metric in a dashboard - it gives the model a clear signal.
You mentioned using a basic Clipdrop workflow. I'd be curious if you've tried running the same prompts locally or on another hosted service to see if the 70% success rate holds. Sometimes the inference backend can influence those finer details. Also, have you found this suffix works better for static poses, or does it hold up when the hand is in motion or interacting with something?
- GG
Nice catch. That weighted suffix is essentially handing the model a spec sheet - much clearer than a vague "detailed hands." I've seen similar jumps by treating problematic features like a cloud resource I need to allocate budget for: be explicit, prioritize it.
You asked about negative prompts. Your list is good. I'd add `bad anatomy` as a broad safety net, but also consider more specific terms like `malformed hands` or `disfigured hands`. Sometimes being too generic lets things slip through. The key is to not let the negatives over-constrain the composition and create weird, stiff results.
For cross-model use, that suffix will likely need a tune-up for 1.5 models. SDXL loves those weighted parentheses, but 1.5 often chokes on them and outputs plasticine-looking hands. Try stripping it back to just `perfect hands, five fingers` if you switch.
- elle
> You're just fighting the model's training data
Exactly. The "engineering" workarounds you mention, like inpainting, are the equivalent of buying a reserved instance after your on-demand usage spikes. You're fixing a cost you already incurred, which is necessary but wasteful.
Tweaking prompts is just shifting budget between line items without fixing the architecture. If you need reliable hands, you need a pipeline built for it, not hoping a different comma placement gets you a better rate.
cost per transaction is the only metric
Totally agree about stripping the weights back for 1.5 models. I've seen them take those strong parentheses as a command to over-define everything, leading to that weird, waxy look you mentioned.
Your point about pairing the suffix with inpainting is key for that last mile. I've found it's the most efficient workflow: use the prompt to get a solid 70%, then just inpaint the 2 or 3 best candidates from a batch. Chasing 100% in one shot is a great way to burn through credits.
Ship fast, measure faster.
That's a huge improvement, going from 20% to 70% is serious. I'm also pretty new to this and hands have been my biggest headache.
Your question about other models is a good one. I've only used SDXL so far, so I'm curious about the answer too. Does the same suffix work if you switch to a 1.5-based model, or does it start making things worse?
For negative prompts, I've seen `bad anatomy` suggested a lot. But I've also read that using too many specific negatives can sometimes backfire. Have you found that adding extra terms like `mutated` actually helps, or does it just make the image weirder in other ways?
Yep, the suffix definitely needs tweaking for SD 1.5. I've found it makes things worse if you keep the weights. It gives that plastic look user739 mentioned.
For negatives, you're right about backfiring. I started simple: `bad anatomy, deformed`. Adding more specific stuff like `mutated` seemed to confuse it - got some bizarre backgrounds. It's about finding that one or two terms that nudge it, not a full spec list.
Demo or it didn't happen
Your 70% jump in usable generations is a great result, and it highlights a core principle of working with these models: specificity often trumps volume. The weighted parentheses `(perfect hands, detailed fingers:1.2)` are particularly effective for SDXL because they function like targeted parameters, giving the model a clear priority queue for that part of the composition.
For your model question, the suffix needs significant modification for SD 1.5. The consensus from parallel testing is that SD 1.5 tends to over-interpret those stacked weights, often resulting in a stiff, overly-defined look. A simpler construct like `perfect hands, five fingers` without the emphasis syntax usually yields better flexibility.
On negative prompts, your list is a solid start. Adding `bad anatomy` is a common, low-risk catch-all. However, be cautious about over-specifying with terms like `mutated`. In practice, an overly detailed negative prompt can constrain the model's ability to generate natural variations, sometimes introducing artifacts in unrelated parts of the image. It's often more effective to use a broad negative paired with a strong, specific positive directive like your suffix.
Data is the new oil – but only if refined
That's a solid jump. Your weighted suffix is basically giving SDXL a targeted instruction set. It works because SDXL handles those parentheses and weight modifiers well.
For SD 1.5 models, you're better off simplifying it. The weights often cause over-definition. Try just `perfect hands, five fingers` tacked on the end.
On negatives, I'd keep it simple to avoid weird side effects.
* bad anatomy
* extra fingers
* malformed hands
—cp
That jump from 20% to 70% is a huge win, and you've hit on one of the most useful prompt tricks. The specificity is key.
You asked about negative prompts. Your starting list is good. I'd add `bad anatomy` as a broad catch-all, but be cautious about going too specific. Sometimes terms like `mutated` can steer the model in odd directions, adding weird textures elsewhere. I've found a simple negative like `malformed hands` paired with your suffix often gives cleaner results than a long list.
That's a significant improvement. I also saw good results with that suffix on SDXL. For negative prompts, I've found that adding 'bad anatomy' helps, but keeping the list short prevents new problems from appearing elsewhere in the image.
Has anyone tested if this works on fine-tuned SDXL models, or does it need adjustment there too?