[Issue]: OpenVINO backend unstable

[alexdsh]

Feature description

Hello! I'm using a 1240-based laptop. I was able to use the GPU with the OpenWino backend (it only accelerates Unet rendering). As far as I understand, this feature is disabled in the latest builds. You're only focusing on Arc versions of Intel GPUs. Is it possible to add compatibility with older Intel GPUs like the Iris XE and Ultra HD? Not everyone has access to an Nvidia or even AMD accelerator. Please add this feature. Only the 2025 commit works more or less stably for me.

Version Platform Description

Имя устройства EclipseX360
Процессор 12th Gen Intel(R) Core(TM) i5-1240P (1.70 GHz)
Оперативная память 16,0 ГБ (доступно: 15,6 ГБ)
Тип системы 64-разрядная операционная система, процессор x64

[RoyalXXX]

@alexdsh I also have an Intel Iris Xe GPU, and I don’t really understand what problem you’re running into. I installed the latest version of SD.Next (updated: 2026-02-07; hash: c9e21a5), and OpenVINO (2025.4.1) works perfectly. I tested it with both SD and SDXL models - image generation is fast and runs on the GPU.

[vladmandic]

@alexdsh like @RoyalXXX said, openvino is still fully supported - what is the actual problem?

[RoyalXXX]

@vladmandic when generating images using the SDXL model with OpenVINO, I noticed a strange behavior. Normally, only the first generation should take a long time because the model is being compiled before the very first run. If the generation settings remain unchanged, all subsequent generations should be very fast since the already compiled model is reused.
However, in SD.Next there is some odd behavior: not only the first generation is slow, but the second one also takes a long time. Only starting from the third generation does it become really fast — the process starts immediately without any noticeable preliminary steps.
It seems that some unnecessary or additional operations are happening during the second generation.

[vladmandic]

@RoyalXXX that sounds unrelated to this issue - can you please create a separate issue and upload a full log so it covers startup/first-generate/second-generate/third-generate

[vladmandic]

@alexdsh please update this issue?

[alexdsh]

@alexdshПожалуйста, обновите информацию по этому вопросу.

Sorry, I was sure the new builds didn't support OpenWino, but after downloading the master version, I tried the OpenWino backend. It works, but it's sad that it only accelerates Unet. The rest of the pipeline runs on the CPU. However, the first launch takes a long time, and the VAE stage is extremely unstable, often throwing errors. I used Grok for installation, and he assured me it was a hardware issue, but I'm glad it works at least somehow. Regarding the latest version, Grok misled me, assuring me that the new builds are only for the ARC architecture and Intel XPU. That's why I wrote to the author asking for extended support.

[alexdsh]

21:40:04-902009 DEBUG Reference: model="black-forest-labs/FLUX.2-klein-base-4B" {'path':
'black-forest-labs/FLUX.2-klein-base-4B', 'preview':
'black-forest-labs--FLUX.2-klein-base-4B.jpg', 'skip': True, 'extras': 'sampler: Default,
cfg_scale: 4.0, steps: 50', 'size': 8.5, 'date': '2025 January'}
21:40:04-907825 DEBUG Search model: name="black-forest-labs/FLUX.2-klein-base-4B"
matched="black-forest-labs/FLUX.2-klein-base-4B" type=reference
21:40:04-907825 INFO Load model: select="black-forest-labs/FLUX.2-klein-base-4B"
21:40:04-921341 DEBUG Cache clear
21:40:04-925350 DEBUG Current model: {'ram': {'total': 15.59, 'rss': 4.76, 'used': 9.89, 'free': 5.7, 'avail': 5.7,
'buffers': 0, 'cached': 0}, 'gpu': {'total': 0, 'used': 0, 'error': 'Torch not compiled with
CUDA enabled', 'swap': 0}, 'job': 'Load model'}
21:40:05-958554 DEBUG GC: current={'gpu': 0, 'ram': 9.88, 'oom': 0} prev={'gpu': 0, 'ram': 9.88} load={'gpu': 0,
'ram': 63} gc={'gpu': 0, 'py': 978} fn=reload_model_weights:unload_model_weights why=unload
time=0.67
21:40:05-966559 DEBUG Unload model: {'ram': {'total': 15.59, 'rss': 4.76, 'used': 9.88, 'free': 5.7, 'avail': 5.7,
'buffers': 0, 'cached': 0}, 'gpu': {'total': 0, 'used': 0, 'error': 'Torch not compiled with
CUDA enabled', 'swap': 0}, 'job': 'Load model'} fn=reference_submit:reload_model_weights
21:40:05-991088 INFO Autodetect model: detect="FLUX2 Klein" class=Flux2KleinPipeline
file="black-forest-labs/FLUX.2-klein-base-4B"
21:40:05-996094 DEBUG Cache clear
21:40:06-021506 DEBUG HF login: no token provided
21:40:06-472376 DEBUG Load model: type=Flux2Klein repo="black-forest-labs/FLUX.2-klein-base-4B"
config={'low_cpu_mem_usage': True, 'torch_dtype': torch.float32, 'load_connected_pipeline':
True, 'safety_checker': None, 'requires_safety_checker': False} offload=balanced
dtype=torch.float32 args={'torch_dtype': torch.float32}
21:40:06-481002 DEBUG Load model: transformer="black-forest-labs/FLUX.2-klein-base-4B" cls=Flux2Transformer2DModel
subfolder=transformer quant="None" loader=default args={'torch_dtype': torch.float32}
21:40:37-904919 TRACE Server: alive=True requests=36 memory=15.59/15.59 status='running' task='Load DiT'
timestamp='20260329214006' current='' id='84e3f03cc05a45e' job=0 jobs=1 total=4 step=0 steps=0
queued=0 uptime=285 elapsed=31.05 eta=None progress=0
21:40:43-211152 DEBUG Load model: text_encoder="Tongyi-MAI/Z-Image-Turbo" cls=Qwen3ForCausalLM quant="None"
loader=default shared=True
Progress 17.32s/it ██████████████ 100% 3/3 00:51 00:00 Loading checkpoint shards
generation_config.json: 100%|██████████████████████████████████████████████████████████| 239/239 [00:00<00:00, 242kB/s]
model_index.json: 100%|███████████████████████████████████████████████████████████████████████| 422/422 [00:00<?, ?B/s]
tokenizer_config.json: 5.40kB [00:00, 753kB/s] 0% 0/13 00:00 ? Fetching 13 files
config.json: 1.54kB [00:00, 174kB/s] ?B/s]
scheduler_config.json: 100%|██████████████████████████████████████████████████████████| 486/486 [00:00<00:00, 60.5kB/s]
chat_template.jinja: 4.17kB [00:00, 2.08MB/s] | 0.00/486 [00:00<?, ?B/s]
special_tokens_map.json: 100%|█████████████████████████████████████████████████████████| 613/613 [00:00<00:00, 299kB/s]
added_tokens.json: 100%|██████████████████████████████████████████████████████████████████████| 707/707 [00:00<?, ?B/s]
merges.txt: 1.67MB [00:00, 5.20MB/s] | 0.00/613 [00:00<?, ?B/s]
config.json: 100%|████████████████████████████████████████████████████████████████████| 531/531 [00:00<00:00, 1.00MB/s]
config.json: 100%|████████████████████████████████████████████████████████████████████████████| 821/821 [00:00<?, ?B/s]
vocab.json: 2.78MB [00:00, 7.81MB/s] | 0.00/11.4M [00:00<?, ?B/s]
tokenizer/tokenizer.json: 100%|███████████████████████████████████████████████████| 11.4M/11.4M [00:03<00:00, 3.71MB/s]
vae/diffusion_pytorch_model.safetensors: 100%|██████████████████████████████████████| 168M/168M [00:37<00:00, 4.45MB/s]
Progress 3.03s/it ████████████████████ 100% 13/13 00:39 00:00 Fetching 13 files██| 11.4M/11.4M [00:03<00:00, 57.2MB/s]
Progress 2.40it/s █████████ 100% 5/5 00:02 00:00 Loading pipeline components... | 0.00/168M [00:02<?, ?B/s]
21:42:25-345611 DEBUG GC: current={'gpu': 0, 'ram': 7.17, 'oom': 0} prev={'gpu': 0, 'ram': 6.65} load={'gpu': 0,
'ram': 46} gc={'gpu': 0, 'py': 17763} fn=load_diffuser_force:load_flux2_klein why=load
time=4.11
21:42:25-393241 DEBUG Setting model: component=vae {'no-half': True, 'slicing': True, 'tiling': False}
21:42:25-396910 DEBUG Setting model: attention="Scaled-Dot-Product"
21:42:25-400534 INFO Offload: type=balanced op=init watermark=0-0.6 gpu=0.00-9.60:16.00 cpu=16.000 limit=0.00
always=[] never=[] pre=True streams=False
21:42:25-427140 DEBUG Module: name=text_encoder cls=Qwen3ForCausalLM size=14.985 params=4022468096 quant=None
21:42:25-432674 DEBUG Module: name=transformer cls=Flux2Transformer2DModel size=14.438 params=3875544576 quant=None
21:42:25-442138 DEBUG Module: name=vae cls=AutoencoderKLFlux2 size=0.313 params=84046115 quant=None
21:42:25-447176 INFO Model class=Flux2KleinPipeline modules=3 size=29.735
21:42:25-462546 INFO Load model: family=f2 time={'total': 139.47, 'load': 139.39} native=1024 memory={'ram':
{'total': 15.59, 'rss': 3.01, 'used': 7.16, 'free': 8.42, 'avail': 8.42, 'buffers': 0,
'cached': 0}, 'gpu': {'total': 0, 'used': 0, 'error': 'Torch not compiled with CUDA enabled',
'swap': 0}, 'job': ''}
21:42:37-423895 TRACE Server: alive=True requests=58 memory=7.14/15.59 status='idle' task='' timestamp=None
current='' id='' job=0 jobs=0 total=5 step=0 steps=0 queued=0 uptime=405 elapsed=11.95 eta=None
progress=0
21:44:38-288352 TRACE Server: alive=True requests=80 memory=8.41/15.59 status='idle' task='' timestamp=None
current='' id='' job=0 jobs=0 total=5 step=0 steps=0 queued=0 uptime=526 elapsed=132.82
eta=None progress=0
21:46:39-184763 TRACE Server: alive=True requests=99 memory=8.53/15.59 status='idle' task='' timestamp=None
current='' id='' job=0 jobs=0 total=5 step=0 steps=0 queued=0 uptime=647 elapsed=253.71
eta=None progress=0
21:46:55-071244 DEBUG Pipeline class change: original=Flux2KleinPipeline target=Flux2KleinPipeline device=cpu
fn=process_images_inner:init
21:46:55-181307 DEBUG Pipeline class change: original=Flux2KleinPipeline target=Flux2KleinPipeline device=cpu
fn=process_images_inner:process_diffusers
21:46:55-283833 WARNING Sampler: "DPM++ 2M" does not accept sigmas
21:46:55-288390 DEBUG Sampler: "Default" cls=FlowMatchEulerDiscreteScheduler config={'num_train_timesteps': 1000,
'shift': 3.0, 'use_dynamic_shifting': True, 'base_shift': 0.5, 'max_shift': 1.15,
'base_image_seq_len': 256, 'max_image_seq_len': 4096, 'invert_sigmas': False, 'shift_terminal':
None, 'use_karras_sigmas': False, 'use_exponential_sigmas': False, 'use_beta_sigmas': False,
'time_shift_type': 'exponential', 'stochastic_sampling': False}
21:46:55-420169 INFO Processing modifiers: apply
21:46:56-188909 INFO Base: pipeline=Flux2KleinPipeline task=IMAGE_2_IMAGE batch=1/1x1 set={'prompt': 1,
'guidance_scale': 6, 'generator': 'cpu:[1333597339]', 'num_inference_steps': 50, 'output_type':
'np', 'width': 512, 'height': 512, 'parser': 'fixed'}
21:47:35-766593 INFO API user=None code=200 http/1.1 GET /sdapi/v1/network/desc 127.0.0.1 0.0631
21:48:39-278306 TRACE Server: alive=True requests=143 memory=15.32/15.59 status='running' task='Inference'
timestamp='20260329214656' current='task(3jnim7ny3ih8h83)' id='69c6283b70534d8' job=0 jobs=1
total=11 step=0 steps=0 queued=0 uptime=767 elapsed=103.04 eta=None progress=0
Progress ?it/s 0% 0/50 00:00 ? Base21:50:39-365847 TRACE Server: alive=True requests=171 memory=15.3/15.59 status='running' task='Inference'
timestamp='20260329214656' current='task(3jnim7ny3ih8h83)' id='69c6283b70534d8' job=0 jobs=1
total=11 step=0 steps=0 queued=0 uptime=887 elapsed=223.11 eta=None progress=0
21:52:01-758010 INFO API user=None code=200 http/1.1 GET /sdapi/v1/network/desc 127.0.0.1 0.2169
21:52:39-647213 TRACE Server: alive=True requests=196 memory=14.67/15.59 status='running' task='Inference'
timestamp='20260329214656' current='task(3jnim7ny3ih8h83)' id='69c6283b70534d8' job=0 jobs=1
total=11 step=0 steps=0 queued=0 uptime=1007 elapsed=343.42 eta=None progress=0
Progress 224.88s/it █▏ 4% 2/50 07:27 2:59:54 Base21:57:31-520489 DEBUG State: interrupt requested

Subject: Issues with Flux.2 Klein and Flux models on Intel Iris Xe iGPU
Hi,
First of all, thank you for the great project! I’m actively using SD.Next on a laptop with Intel Core i5-1240P + Iris Xe Graphics (16 GB RAM).
I wanted to report some problems with Flux model support:

Flux.1-schnell / dev sometimes runs on GPU, but after several restarts or settings changes, it often falls back to CPU and stops detecting the safetensors file if it’s placed in models/Stable-diffusion. I have to manually create the folder structure models/diffusers/models--black-forest-labs--FLUX.1-schnell.
Flux.2 Klein (4B distilled fp8) doesn’t want to run on GPU at all — it immediately falls back to CPU, even though the model is relatively light (~4.2 GB). The DiT/U-Net part barely uses the iGPU.
Similar behavior with LTX-Video (I2V) — the model is visible, but runs only on CPU.

It would be great to improve Flux family support for integrated Intel graphics (Iris Xe / UHD). Right now, for comfortable work with Flux I have to use ComfyUI, even though SD.Next has a much more convenient interface.
If needed, I can provide logs, config files, and exact model names where the issue occurs.
Thank you for your work!

[vladmandic]

dev branch has been updated with significant updates to openvino processing.
if you can try and report any updates from your side?