Inference crashes on CUDA out of memory
A pipeline that worked on a 24 GB card is being run on a 16 GB one, or a LoRA stack and ControlNet together are pushing past the memory budget without any offload configured.
Pipeline fits the GPU in use
Running Stable Diffusion?
A notebook can generate an image in ComfyUI on a borrowed GPU; shipping that same output as a feature inside your product, running on predictable hardware, at a price you can show finance, is a different piece of work. We take a weekend prototype and turn it into something queueable, observable and safe to point at real users.
What it is
Stable Diffusion is an open weights latent diffusion image model first released by Stability AI and the CompVis group in August 2022. It generates images from text prompts, image prompts, structural guides via ControlNet, and reference images via IP Adapter, all on consumer GPU hardware.
The ecosystem now includes SDXL, SD3 and SD3.5, SDXL Turbo for near real time generation, FLUX models from Black Forest Labs that run on the same tooling, and thousands of community LoRAs on Civitai and Hugging Face. Production use typically runs through the diffusers library, ComfyUI or Automatic1111, deployed behind a queue on dedicated GPU workers or a serverless GPU provider.
The honest version
No technology is all upside. This is where Stable Diffusion earns its place and where it tends to cause trouble.
Sound familiar?
The Stable Diffusion problems that bring people to us, and what you have once each one is fixed.
A pipeline that worked on a 24 GB card is being run on a 16 GB one, or a LoRA stack and ControlNet together are pushing past the memory budget without any offload configured.
Pipeline fits the GPU in use
A VAE or LoRA trained for one resolution being fed latents from a pipeline configured for another, which crashes somewhere mid denoise rather than at load time.
Components aligned to one resolution
A LoRA trained against SD 1.5 being loaded into an SDXL pipeline, or a community LoRA that was quietly repacked and now has mismatched layer names.
LoRA matched to base, loader fixed
Full precision SDXL at thirty steps with a refiner and ControlNet is simply slow on the current hardware. We swap in a turbo variant, reduce steps and fuse components to hit the budget.
Target latency met under load
The bundled checker fires on brand imagery or product shots because the training distribution flagged a texture or skin tone, leaving users with an opaque failure.
Classifier replaced with usable policy
Each request on an idle worker loads a checkpoint, a LoRA stack and ControlNet weights from scratch. Users see a half minute delay before any pixels appear.
Warm pool with real latency floor
A truncated download, a wrong format for the diffusers version, or a safetensors file that was quietly renamed from a different architecture altogether.
Checkpoints verified and pinned
Dedicated GPUs idling overnight, spot instances recreated hourly, or a workflow running every step at maximum precision when a lower one would be indistinguishable.
Right sized to actual traffic
What we do
Turn one ComfyUI workflow into a service, or let us run the whole image pipeline.
How it works
Tell us what's going wrong. We look at it, then tell you plainly what's wrong and what it will cost.
We fix the cause, not the symptom. If something is down, getting it back up comes first.
Backups, monitoring, updates and access, so the same problem can't quietly come back.
You get a plain-language write-up. Or we stay on for hosting, upkeep and new work.
Often found together
We work on the whole stack, so the problem can be wherever it actually is.
The common causes are a pipeline that stacks SDXL, ControlNet, a refiner and a LoRA on a card that cannot hold them all, a batch size too large for the resolution, and VAE operations running at the same time as the UNet instead of sequentially. We add offload, pick a smaller checkpoint when it fits, and fit the memory budget before adding features back in.
Yes. LoRA training on a few dozen to a few hundred curated images produces a reusable weight file, in hours on a single GPU, that adapts the base model to your style. Dreambooth goes further when you want the model to learn a specific subject. We handle the dataset preparation, the training runs and the evaluation against a held out set so you can see where the model is reliable.
Replace the bundled NSFW checker with a classifier you have measured, filter prompts before they reach the model, keep logs of inputs and outputs for audit, and watermark or sign generated images where downstream systems need to tell them apart from real photos. The bundled defaults are not a safety posture; they are a placeholder.
Managed APIs are the right call for low or spiky volume, where you pay per second and someone else deals with GPUs. Self hosted wins when you have steady traffic, IP sensitive content, or very specific model stacks. We will run the numbers against your actual request rate rather than quote a sticker price.
SDXL and newer models produce commercial grade images on a single L4, A10 or RTX 4090 class GPU, and faster on an A100 or H100. Smaller SD 1.5 workloads still run on an 8 GB card at modest resolution. The right hardware depends on the resolution, the step count, how many LoRAs you stack and your latency budget, which is a short conversation before any order.
Estimates are free of charge. Describe the workflow, send a sample output or an error from the logs, and we send a price before any work begins.
Free estimate
Tell us what you're running and what's going wrong. A person reads it and comes back with an estimate.
Your message is with a real person, and we'll get back to you soon.
Rather talk now? Call us: