Image generation runtime

Self-hosting AUTOMATIC1111: feature-frozen at v1.10.1, still the UI every extension was written for

Q: How much VRAM does Automatic1111 need? Can a 4GB card really run it?

The README does say 4GB cards are supported with reports of 2GB working, and that relies on --lowvram ("sacrificing a lot of speed for very low VRAM usage"). The honest numbers come from the official Optimizations wiki: on an RTX 3060 running SD 1.5 with xformers, peak VRAM was 2.8GB at batch 1, 2, 4 and 8 and only 4.1GB at batch 16, while the default Doggettx path needed 6.6GB at batch 8. SDXL is a different ledger — the base weights alone are 6.94GB, an 8GB card only limps along with --medvram-sdxl, and 12GB is the realistic floor. NexGPU's cheapest 24GB card is the RTX 3090 at $0.193/GPU-hr, which covers all three tiers with room to spare.

Q: Can A1111 run Flux? Can it run SD 3.5?

Upstream, neither. v1.10.0 added SD3 Medium, and SD 3.5 arrived three months after master froze, so it never landed; Flux never landed at all. To run either in this familiar UI you need Forge (BitsAndBytes NF4 and GGUF — in city96's quant set, Q4_K_S is 6.81GB, Q8_0 is 12.71GB and F16 is 23.80GB, none of which include T5-XXL) or the newer forge-neo branch. Mapping that to cards: Q4 quants fit an RTX 4090 24GB at $0.540/GPU-hr, while Q8 plus an fp16 T5 wants an RTX 5090 32GB at $0.723 or an RTX A6000 48GB at $0.817 — all available on NexGPU and billed by the second.

Q: Why does AUTOMATIC1111 fail with "no kernel image is available" on an RTX 5090?

Because launch_utils.py defaults to pip install torch==2.1.2 torchvision==0.16.2 --extra-index-url https://download.pytorch.org/whl/cu121, and torch 2.1.2 on cu121 has no Blackwell (sm_120) kernels at all. The matching pin, xformers==0.0.23.post1, has no RTX 50-series build either — forge-neo's README says flatly that xformers does not support RTX 50s. The fix is to override TORCH_COMMAND with a cu128 build before launch and replace --xformers with --opt-sdp-attention ("Enable scaled dot product cross-attention layer optimization; requires PyTorch 2.*"). To skip the whole problem, rent an RTX 3090 24GB at $0.193 or an RTX 4090 24GB at $0.540 on NexGPU, where the stock pins just work and you are generating on first boot.

Q: My fresh install hangs downloading the default model — v1-5-pruned-emaonly won't fetch. What now?

It is not your network. modules/sd_models.py still hard-codes the default download to runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors, and that repo now returns 401. The weights live at stable-diffusion-v1-5/stable-diffusion-v1-5 instead — emaonly is 4.27GB, the full checkpoint 7.70GB. Either wget it into models/Stable-diffusion yourself or pass --no-download-sd-model to skip the auto-download entirely. NexGPU has nodes in 51 countries and regions, so pick a region close to Hugging Face and 20GB of weights land quickly; keep the volume and they are still there next boot, no re-download needed.

Q: What is the safe way to reach the WebUI in the cloud? Should I use --share?

No. --share passes share=True to gradio and exposes the UI through gradio's own site, and casually adding --listen to bind 0.0.0.0:7860 is just as bad, because A1111 has no authentication by default — stack --enable-insecure-extension-access on top and you have effectively handed out remote code execution. Bind to localhost and forward with an SSH tunnel, -L 7860:127.0.0.1:7860. For a backend service, add --api or go headless with --nowebui. Every NexGPU instance ships SSH, Jupyter, a web terminal, a REST API and a CLI, so the tunnel is a two-second job, and support is bilingual over Telegram with no ticket queue.

The master branch stopped at July 2024 while the pins still say torch 2.1.2+cu121, gradio 3.41.2, transformers 4.30.2. This page gives you the real VRAM numbers, the real install traps, and the NexGPU card that matches each one — before you rent anything.

Automatic1111 · self-hosted

AUTOMATIC1111/stable-diffusion-webui is a 164.6k-star, 30.6k-fork AGPL-3.0 project and the de facto interface standard for the entire Stable Diffusion extension ecosystem. Its status needs stating plainly: it is feature-frozen at v1.10.0, the release that brought SD3 support, six new schedulers (Align Your Steps, KL Optimal, Normal, DDIM, Simple, Beta), the DDIM CFG++ sampler, and the option to skip CFG on early generation steps. The only release since, v1.10.1, changed exactly one thing — the changelog reads "fix image upscale on cpu". The last commit on master is dated 2024-07-27. It is not dead, though: the dev branch still takes maintenance fixes — uv on Linux in Oct 2025, a stable_diffusion_repo URL update in Dec 2025, CLIP installation failures in Feb 2026, and a setuptools install fix in Mar 2026. All of it is work to keep the thing installable.

What really argues against installing it on your own machine is the dependency list. requirements_versions.txt still pins gradio==3.41.2, transformers==4.30.2, numpy==1.26.2, Pillow==9.5.0, pytorch_lightning==1.9.4, fastapi==0.94.0 and setuptools==69.5.1, while launch_utils.py defaults to pip install torch==2.1.2 torchvision==0.16.2 against the cu121 index with xformers pinned to 0.0.23.post1. webui.sh looks for python3.10; on Windows only 3.10 is accepted, and anything other than 3.10.6 triggers a warning. numpy locked to 1.x means nothing modern can share that venv. Then there is the trap every fresh install hits: modules/sd_models.py still hard-codes the default checkpoint download to runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors, and that Hugging Face repo now returns 401 — so a clean install with an empty models/Stable-diffusion directory stalls right there.

The forks matter too. lllyasviel's Forge is built on webui 1.10.1 with Gradio 4, and it added what upstream never had: Flux support via BitsAndBytes NF4 and GGUF (Q8_0/Q5_0/Q5_1/Q4_0/Q4_1), multiple ControlNets, IP-Adapter and Instant-ID; its main branch last committed on 2025-06-26, around the time Chroma and fp8-scaled support landed. Haoming02's forge-classic keeps its classic branch as an archive on Gradio 3.41.2 with only SD1 and SDXL, while the active neo branch recommends torch 2.10.0+cu130, xformers 0.0.34 and Python 3.11.9, drops SD2, SD3, hypernetworks, Deepbooru, TI training and bitsandbytes, and adds SageAttention, FlashAttention and torch.compile. Same UI, same extensions, four mutually incompatible dependency stacks — which is precisely why this belongs on a rented box, not your workstation. NexGPU runs 1,175 verified rentable nodes across 51 countries and regions, 2,498 GPUs and 75 GPU models, with AUTOMATIC1111, Stable Diffusion, ComfyUI and PyTorch among 2,000+ prebuilt images, billed per second with no minimum, no setup fee and no quota request. Blowing up an install and starting over costs cents.

01 —

Versions, forks, and what each can actually load

Upstream frozen, dev on life support, Forge carrying the torch — pick the wrong branch and your model simply will not load.

Version	Parameters	VRAM	Context	Notes
v1.10.1 (the only current release)	gradio 3.41.2 · torch 2.1.2+cu121 · xformers 0.0.23.post1	SD 1.5 peaks around 2.8GB with xformers; SDXL practical floor 12GB	SD 1.x / SD 2.x / SDXL / SD3 Medium	AGPL-3.0. Its only delta from 1.10.0 is "fix image upscale on cpu", so functionally this is the frozen build. ControlNet, ADetailer and the rest of the mainstream extensions target this interface, and compatibility here is still the best you will get.
v1.10.0 (the feature freeze point)	Final batch of merges on master, 2024-07-27	SD3 Medium bundles: 4.34GB / 5.97GB / 10.87GB (t5 fp8) / 15.76GB (t5 fp16)	Adds SD3 support, Euler recommended	Brought six schedulers (Align Your Steps, KL Optimal, Normal, DDIM, Simple, Beta), the DDIM CFG++ sampler, skipping CFG on early steps, and hash checking of downloaded models. SD 3.5 shipped three months after this freeze and never made it upstream.
dev branch (still maintained)	Pins nearly identical to master; only safetensors moved 0.4.2 → 0.4.5	Same as v1.10.1	No new model support, just installability	Recent commits are all install repairs: fix uv on linux (Oct 2025), update stable_diffusion_repo url (Dec 2025), Fix CLIP installation failures (Feb 2026), setuptools install fix (Mar 2026). On a fresh box in 2026, clone -b dev saves you an afternoon versus master.
Forge (lllyasviel)	Based on webui 1.10.1 · Gradio 4.40 · recommends CUDA 12.1 + PyTorch 2.3.1	Flux GGUF Q4_K_S 6.81GB / Q8_0 12.71GB / F16 23.80GB (transformer only, no T5-XXL)	SD 1.5 / SDXL / SD3 / Flux	Upstream A1111 never supported Flux for a single day; running Flux in this UI means the Forge line. Supports BitsAndBytes NF4 and GGUF Q8_0/Q5_0/Q5_1/Q4_0/Q4_1, multiple ControlNets, IP-Adapter and Instant-ID. The README's own advice: do not set GPU Weight too high — lowering it solves 99% of problems. Main branch last committed 2025-06-26.
forge-classic / neo (Haoming02)	neo: torch 2.10.0+cu130 · xformers 0.0.34 · Python 3.11.9	Depends on the loaded checkpoint; neo's fp8 and torch.compile flatten the curve	neo tracks the newer open-weight releases	The classic branch is explicitly an archive (Gradio 3.41.2, SD1 and SDXL only, critical fixes only). neo is the live line: it removes SD2, SD3, hypernetworks, CLIP Interrogator, Deepbooru, TI training and bitsandbytes, and adds SageAttention, FlashAttention and torch.compile. Its README states outright that xformers does not support RTX 50s.

02 —

Pick the card by what you will actually load

A1111's VRAM bill has nothing to do with the UI — it is which generation of weights you load, whether xformers is on, and how many checkpoints stay resident.

Day-to-day SD 1.5 / SDXL generation with ControlNet and batch API runs
RTX 3090 24GB$0.193/GPU-hr
Ampere is exactly what A1111's hard-pinned torch 2.1.2+cu121 and xformers 0.0.23.post1 were built for, so --xformers works out of the box. The official wiki measured SD 1.5 on an RTX 3060 peaking at 2.8GB across batch 1 through 8 with xformers and 4.1GB at batch 16, leaving the whole 24GB for SDXL resolution and hires.fix.
Heavy SDXL 1024 output with ADetailer and hires.fix chained together
RTX 4090 24GB$0.540/GPU-hr
Same 24GB, a clear step up in single-card throughput. sd_xl_base_1.0.safetensors at 6.94GB plus the refiner, an fp16 VAE and several resident LoRAs still leaves headroom, so you never have to trade speed away with --medvram-sdxl.
SD3 Medium with the fp16 T5-XXL, or several checkpoints hot-swapping
RTX A6000 48GB$0.817/GPU-hr
sd3_medium_incl_clips_t5xxlfp16.safetensors is 15.76GB in a single file; stack SDXL and a few community models on top and switch between them in settings, and only 48GB avoids a reload every time. It is roughly 25% cheaper than the A100 SXM4 80GB at $1.088.
Switching to Forge / forge-neo for Flux and newer weights
RTX 5090 32GB$0.723/GPU-hr
32GB holds the 12.71GB Q8_0 Flux transformer plus T5-XXL and the VAE. Note that Blackwell needs torch built against CUDA 12.8 or newer — A1111's pinned 2.1.2+cu121 has no kernels for it, so override TORCH_COMMAND and swap --xformers for --opt-sdp-attention.

03 —

Getting it running on a rented card

Four steps, every command copy-pasteable, routing around the 401 on the default checkpoint URL and the open-to-the-internet mistake.

01
Boot the instance and tunnel port 7860 home
Pick a 24GB card in the console and select the AUTOMATIC1111 image, or take Ubuntu CLI / PyTorch and install it yourself. Once it is up, resist reaching for --listen: forward the port over SSH first, because A1111 ships with no authentication of any kind.
```
ssh -N -L 7860:127.0.0.1:7860 -p <port> root@<node-ip>
```
02
Nail Python to 3.10 and clone the dev branch
webui.sh looks for python3.10 and launch_utils.py warns on anything other than 3.10.6; pins like numpy==1.26.2 and Pillow==9.5.0 have no wheels for 3.12 or 3.13. The dev branch carries the 2025-2026 CLIP and setuptools install fixes, so it lands cleanly far more often than master.
```
sudo apt install -y wget git python3.10 python3.10-venv libgl1 libglib2.0-0 && git clone -b dev https://github.com/AUTOMATIC1111/stable-diffusion-webui && cd stable-diffusion-webui
```
03
Put a checkpoint in place before first launch
modules/sd_models.py still points the default download at runwayml/stable-diffusion-v1-5, and that Hugging Face repo now returns 401. The weights moved to stable-diffusion-v1-5/stable-diffusion-v1-5 (emaonly 4.27GB, full 7.70GB). For SDXL, grab sd_xl_base_1.0.safetensors at 6.94GB instead.
```
wget -P models/Stable-diffusion https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors
```
04
Launch with the REST API and wire it into your own service
--api mounts /sdapi/v1/txt2img, /sdapi/v1/img2img, /sdapi/v1/png-info, /sdapi/v1/progress, /sdapi/v1/sd-models and /sdapi/v1/options; if you want the backend only, add --nowebui ("Only launch the API, without the UI"). --no-download-sd-model skips the broken auto-download. On Blackwell cards like the RTX 5090, first export TORCH_COMMAND="pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128" and swap --xformers for --opt-sdp-attention.
```
bash webui.sh -f --xformers --api --no-download-sd-model --port 7860
```

What one full run actually costs

A batch SDXL session on an RTX 3090 24GB: install (apt dependencies, venv, the full pinned pip set) about 15 minutes, or 0.25 hr; pulling sd_xl_base_1.0.safetensors at 6.94GB plus two community checkpoints and a set of ControlNet weights, roughly 20GB in 10 minutes, or 0.17 hr; then 3 hours of actual generation. That is 3.42 hr × $0.193 = $0.660. Provision a 60GB volume: destroy it when you finish and storage stops billing; keep it overnight and the $0.414/GB-month median works out to about $0.0138/GB per day, so 60GB costs $0.828. Downloading 800 PNGs at 1024×1024, about 1.2GB, at the $0.0081/GB median egress adds $0.010. So a complete run is $0.67 without keeping the volume, or $1.50 with it held for a day. Want it faster? RTX 4090 24GB is 3.42 × $0.540 = $1.847. Need 48GB for multiple resident checkpoints? RTX A6000 is 3.42 × $0.817 = $2.794. Billing is metered per second and quoted per hour; compute stops the moment the instance stops, and only the volume keeps accruing until it is destroyed.

04 —

FAQ

Is AUTOMATIC1111 still maintained? Is it worth installing in 2026?

Feature-wise it is frozen: the last commit on master is 2024-07-27, and the only release since, v1.10.1, changed exactly one thing — fixing image upscale on CPU. But the dev branch still takes install fixes, most recently on 2026-03-02 for pip install 'setuptools<70' failing under cmd, preceded by a CLIP installation fix in Feb 2026. The reason to install it is the ecosystem: thousands of extensions, ControlNet, ADetailer and nearly every community tutorial target this UI and this API, and nothing has replaced that. If you need newer models, spin up a second box on the Forge line. On NexGPU both paths take minutes — AUTOMATIC1111 is one of the 2,000+ prebuilt images, billing is per second, and destroying a broken install to start fresh costs almost nothing.

How much VRAM does Automatic1111 need? Can a 4GB card really run it?

The README does say 4GB cards are supported with reports of 2GB working, and that relies on --lowvram ("sacrificing a lot of speed for very low VRAM usage"). The honest numbers come from the official Optimizations wiki: on an RTX 3060 running SD 1.5 with xformers, peak VRAM was 2.8GB at batch 1, 2, 4 and 8 and only 4.1GB at batch 16, while the default Doggettx path needed 6.6GB at batch 8. SDXL is a different ledger — the base weights alone are 6.94GB, an 8GB card only limps along with --medvram-sdxl, and 12GB is the realistic floor. NexGPU's cheapest 24GB card is the RTX 3090 at $0.193/GPU-hr, which covers all three tiers with room to spare.

Can A1111 run Flux? Can it run SD 3.5?

Upstream, neither. v1.10.0 added SD3 Medium, and SD 3.5 arrived three months after master froze, so it never landed; Flux never landed at all. To run either in this familiar UI you need Forge (BitsAndBytes NF4 and GGUF — in city96's quant set, Q4_K_S is 6.81GB, Q8_0 is 12.71GB and F16 is 23.80GB, none of which include T5-XXL) or the newer forge-neo branch. Mapping that to cards: Q4 quants fit an RTX 4090 24GB at $0.540/GPU-hr, while Q8 plus an fp16 T5 wants an RTX 5090 32GB at $0.723 or an RTX A6000 48GB at $0.817 — all available on NexGPU and billed by the second.

Why does AUTOMATIC1111 fail with "no kernel image is available" on an RTX 5090?

Because launch_utils.py defaults to pip install torch==2.1.2 torchvision==0.16.2 --extra-index-url https://download.pytorch.org/whl/cu121, and torch 2.1.2 on cu121 has no Blackwell (sm_120) kernels at all. The matching pin, xformers==0.0.23.post1, has no RTX 50-series build either — forge-neo's README says flatly that xformers does not support RTX 50s. The fix is to override TORCH_COMMAND with a cu128 build before launch and replace --xformers with --opt-sdp-attention ("Enable scaled dot product cross-attention layer optimization; requires PyTorch 2.*"). To skip the whole problem, rent an RTX 3090 24GB at $0.193 or an RTX 4090 24GB at $0.540 on NexGPU, where the stock pins just work and you are generating on first boot.

My fresh install hangs downloading the default model — v1-5-pruned-emaonly won't fetch. What now?

It is not your network. modules/sd_models.py still hard-codes the default download to runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors, and that repo now returns 401. The weights live at stable-diffusion-v1-5/stable-diffusion-v1-5 instead — emaonly is 4.27GB, the full checkpoint 7.70GB. Either wget it into models/Stable-diffusion yourself or pass --no-download-sd-model to skip the auto-download entirely. NexGPU has nodes in 51 countries and regions, so pick a region close to Hugging Face and 20GB of weights land quickly; keep the volume and they are still there next boot, no re-download needed.

What is the safe way to reach the WebUI in the cloud? Should I use --share?

No. --share passes share=True to gradio and exposes the UI through gradio's own site, and casually adding --listen to bind 0.0.0.0:7860 is just as bad, because A1111 has no authentication by default — stack --enable-insecure-extension-access on top and you have effectively handed out remote code execution. Bind to localhost and forward with an SSH tunnel, -L 7860:127.0.0.1:7860. For a backend service, add --api or go headless with --nowebui. Every NexGPU instance ships SSH, Jupyter, a web terminal, a REST API and a CLI, so the tunnel is a two-second job, and support is bilingual over Telegram with no ticket queue.

Start building on NexGPU

Enterprise R&D team or solo developer — either way, your first job can be running in minutes.