Cited, not measured on the local desktop
Local image generation means a diffusion checkpoint runs on your own GPU and writes a picture from a prompt — Stable Diffusion, SDXL, or Flux-class families — without sending the job to a hosted API. On AMDai the first engine is ComfyUI: a node graph on HIP/ROCm PyTorch. It already runs on the local desktop R9700 box. It is not llama.cpp Vulkan, and it does not speak llama-server or Strata /v1. llama.cpp stays first for local LLMs.
How we test — Measured cells cite official amdai-bench/run@1 cards. Estimates stay labeled. Q3 is estimate-only. Q1/Q2 stay off.
You download a diffusion checkpoint, load it in a graph, type a prompt, and the GPU denoises latents into a picture. That is a different job from a local Qwen chat. Weights are not GGUF Q4–Q6 files. The shopper question is which family fits the VRAM sticker, not which tok/s you will see — we have no image-gen run card yet.
Parameter count and dtype come from the publisher card. Fit on a Radeon is an estimate unless a cell says measured. None does. We do not invent a VRAM megabyte figure.
0.9B params (F32 on the card)
Official card lists 0.9B parameters and names the ema-only safetensors as using less VRAM for inference. Training resolution on the card is 512×512. No published VRAM-gigabyte figure.
3B params (F32 on the card; fp16 variant documented)
Official card lists 3B parameters and documents pipe.enable_model_cpu_offload when GPU VRAM is limited. No published VRAM-gigabyte figure.
12B params (BF16 on the card)
Official card is a 12 billion parameter rectified-flow transformer and documents enable_model_cpu_offload to save VRAM. The same card names ComfyUI for local inference. No published VRAM-gigabyte figure.
Classes use AMD-published VRAM stickers from the GPU directory. Every fit cell is an estimate. Every speed cell is not yet measured by AMDai.
| VRAM class | Directory pages | Fit (estimate) | Speed |
|---|---|---|---|
| 32 GB discrete | Estimate. Most discrete headroom on this list for SD 1.5, SDXL, and Flux-class checkpoints. Publisher cards do not publish a VRAM-gigabyte number; resolution, batch, VAE, and attention still decide a run. | not yet measured by AMDai | |
| 24 GB discrete | Estimate. More VRAM than the 20 / 16 / 12 GB classes. SD 1.5 and SDXL are the more likely families; a 12B BF16 Flux-class file is a larger checkpoint than 3B SDXL — still an estimate. | not yet measured by AMDai | |
| 20 GB discrete | Estimate. Between the 24 GB and 16 GB classes. Same families; Flux-class is tighter than SDXL. Confirm the 20 GB sticker before you assume a 24 GB fit. | not yet measured by AMDai | |
| 16 GB discrete | Estimate. SD 1.5 is the more likely family. SDXL may need a lower resolution or the CPU-offload path the SDXL card documents. Flux-class 12B BF16 is tighter on 16 GB than on 24 / 32 GB. | not yet measured by AMDai | |
| 12 GB discrete | Estimate. SD 1.5 is the more likely family. SDXL is tighter. Flux-class is unlikely without the offload the Flux card documents. Confirm the 12 GB sticker. | not yet measured by AMDai | |
| 8 GB discrete | Estimate. 8 GB is tighter than the 12 / 16 GB sheets. SD 1.5 is the more likely family. Confirm the 8 GB sticker — this is not the 16 GB RX 7600 XT. | not yet measured by AMDai | |
| Shared memory (Strix Halo) | Estimate. This APU drinks from the same DIMMs the OS uses — not a discrete VRAM sticker. Size the kit on the Strix Halo sheet. Do not treat it as a 32 GB card. | not yet measured by AMDai |
1 · Cited, not measured on the local desktop
Bare-metal Ubuntu 24.04+ is the AMDai measured path for ROCm on the local desktop. Image-gen follows that OS even though we have no first-party image-gen card yet. See the GPU in rocminfo before you download a checkpoint. Windows is a cited second path, not the first try.
2 · Cited, not measured on the local desktop
ComfyUI needs a ROCm-enabled PyTorch in the same Python environment. AMD’s Install PyTorch for ROCm page documents wheels and Docker. PyTorch’s Get started locally matrix is the official wheel picker: choose Linux, Python, pip, and ROCm. Instinct-oriented Docker tags are not R9700 pins. We do not invent a wheel version.
3 · Cited, not measured on the local desktop
AMD’s ComfyUI page: clone https://github.com/comfyanonymous/ComfyUI.git, activate the ROCm PyTorch venv, then pip install -r ComfyUI/requirements.txt. Start with python ComfyUI/main.py. The official prompt is To see the GUI go to: http://127.0.0.1:8188. Do not treat Instinct Docker tags as a R9700 pin.
4 · Cited, not measured on the local desktop
VRAM limits which family you try first. Cite the publisher card for parameter count and dtype. Fit on a given SKU is an estimate unless AMDai measured it — we have not. Start smaller (SD 1.5) on 12 / 16 GB class cards. The 32 GB R9700 has the most discrete headroom on this list. Seconds-per-image stay not yet measured by AMDai.
Only steps AMD, ComfyUI, or PyTorch document. Everything else is marked unofficial.
AMD’s ComfyUI page documents a Windows install after ROCm-enabled PyTorch: pip install -r ComfyUI\requirements.txt, then python ComfyUI\main.py, same 127.0.0.1:8188 GUI. That page also warns ComfyUI might not start if Smart App Control is enabled. HIP SDK for Windows is AMD’s documented Windows runtime subset. The Radeon limitations page is Windows PyTorch-only, with WSL caveats. Cited, not measured on the local desktop.
PyTorch’s official Get started locally matrix is the wheel picker AMDai cites. The ROCm wheel path AMD documents is Linux + pip + ROCm. Do not invent a Windows ROCm pip pin. Ubuntu 24.04+ on bare metal is still the path that matches local desktop. WSL2 ROCm is a technical preview on the 10.1 notes — not a measured image-gen path.
Community Windows wrappers, DirectML forks, and translation layers that AMD, ComfyUI, or PyTorch do not document stay unofficial. AMDai does not walk them and does not invent a Windows seconds-per-image. If the official AMD page does not list the step, stop.
A local LLM on llama.cpp Vulkan and ComfyUI on HIP/ROCm PyTorch can share a box. They do not share a /v1. On one card, VRAM is one pool — do not invent a split. Size the LLM in the calculator; image-gen VRAM stays an estimate. Minimum system RAM needed and RAM needed at full context are the shopper RAM lines for the LLM. A 64 GB kit estimate is the consumer shopper default (8 GB OS/desktop plus 16 GB headroom plus whatever the LLM offload and PyTorch need) — labeled estimate, not a measured ComfyUI peak. On a second card, read /multi-gpu. We have not measured image-gen-plus-LLM on the local desktop.
Shopper RAM labels stay Minimum system RAM needed and RAM needed at full context for the LLM side. Open the calculator for that load. A second card is a /multi-gpu question, not an invented VRAM split.
ComfyUI is HIP/ROCm PyTorch. If the process never sees /dev/kfd, or PyTorch reports no GPU, you are on CPU. Check groups, the device nodes, and the wheel’s gfx list. This is not a llama.cpp Vulkan miss.
A checkpoint that loads can still OOM at a high resolution or batch. Publisher cards document CPU offload for SDXL and Flux when VRAM is limited. We do not invent a megabyte figure. Lower resolution or batch first. If a local LLM shares the same card, read /multi-gpu and the calculator — do not stack two full VRAM loads and guess.
AMD’s PyTorch install page documents hipErrorNoBinaryForGpu when the wheel or build omits the current gfx. HSA_OVERRIDE_GFX_VERSION is an unofficial way to boot a binary on a gfx the build did not list — a research path, not a recommendation. Current gfx1201 cards should not need it on ROCm 10.x. Confirm rocminfo before you override.