Update 2026-07-05: Krea 2 Turbo v2.0 workflow added
PhotoFlow Krea 2 v2.0
Upscaling has been improved in version 2.0. There are now presets for the settings (sampler, scheduler, and denoise settings) . This allows upscaling to be performed in pixel space to achieve upscales with minimal denoising, resulting in smoother backgrounds that closely resemble the original image. In addition, the euler_ancestral_cfg_pp sampler in combination with the linear_quadratic scheduler is now recommended for high image quality. For fast generation, euler_ancestral with simple can still be used. There are also now filters to correct the colors of the final image, which can become flat due to repeated sampling.
⚠️ Note that the latest version 0.27.0 of ComfyUI (July 2027) is required to use the fast int8 models.
Update 2026-07-02: Krea 2 Turbo v1.0 workflow added
PhotoFlow Krea 2 v1.0
Krea 2 Turbo is the new king of realism and is suitable for both SFW and NSFW content. This workflow is simple, super fast, and optimized for realism.
⚠️ Note that the latest version 0.27.0 of ComfyUI (July 2027) is required to use the fast int8 models.
Update 2026-03-30: Qwen-Image-2512 Nunchaku v1.0 workflow added
PhotoFlow Qwen-Image-2512 Nunchaku v1.0
Qwen-Image-2512 offers exceptional prompt comprehension and delivers outstanding output quality. However, this comes at the cost of very long generation times due to the model's size and the large number of processing steps required. A practical solution in this workflow is to use a heavily quantized Nunchaku FP4 version, which maintains surprisingly high quality. When combined with an 8-step Lightning LoRA, generation speed improves dramatically while image quality remains excellent.
⚠️ Nunchaku requires a few additional setup steps. Please follow the instructions provided in the workflow.
Update 2026-03-21: Flux.2 klein 9B v1.1 workflow added
PhotoFlow Flux.2 klein 9B v1.1
Version 1.1 is simply a hotfix that resolves an issue with an incorrect calculation of the tile size in the Refiner step. Additionally, the EasyLoraStack has been replaced by the Power Lora Loader. The group bypass for all additional steps now also takes the previews into account accordingly.
Update 2026-03-04: Flux.2 klein 9B v1.0 workflow added
PhotoFlow Flux.2 klein 9B v1.0
Flux.2 klein 9B is a fast, 4‑step model supporting image generation up to 4 MP. The workflow starts with a recommended 2 MP base image, followed by optional face refinement and SeedVR2 upscaling to 5 MP for added fine detail. A subsequent refinement step removes SeedVR2 artefacts and enables optional upscaling up to 16 MP. A subtle touch of film grain completes the process, enhancing realism.
Update 2025-12-07: Z-Image-Turbo v2.0 workflow added
PhotoFlow Z-Image-Turbo v2.0
Version 2.0 introduces an optional Refiner and FaceDetailer, along with a recommended set of LoRAs designed to enhance photorealism and overall image quality.
⚠️⚠️⚠️
Make sure you're using ComfyUI version 0.3.76 or newer, since earlier versions contained a bug that affected the correct application of Z-Image-Turbo LoRAs.
Also ensure your SeedVR2 node is up to date, as a previous issue causing visual artifacts has been fixed.
⚠️⚠️⚠️
♥️♥️♥️
A huge thank you to all the creators of the LoRAs recommended in version 2.0. Your work makes this level of quality possible!
All images in the gallery were generated using the same set of LoRAs with identical strength settings, so in most cases no additional tweaking is required 🙂.
Special thanks to Danrisi (NiceGirls UltraReal, Lenovo UltraReal), aip0pp (PhotoReal BetterNudes), HackAfterDark (After Dark), DirtMcGirt (psxZStyle - Photography Style Lora), ConflictX (35mm Photo), and oleshevakatya (Movie Wide Angle).
♥️♥️♥️
Update 2025-11-29: Z-Image-Turbo workflow added
PhotoFlow Z-Image-Turbo
Z-Index Turbo is super fast, can generate images in 2 megapixel resolution without upscaling, and focuses on realism, so no LoRAs are required.
Update 2025-11-23: Qwen workflow added
PhotoFlow Qwen
Until now, I had been hesitant to create a workflow for Qwen because, although its image quality is outstanding, the generation times can be very long when aiming for maximum quality. The Lightning LoRA helped reduce waiting times significantly, but using it alone led to a quality drop that didn’t meet my standards. However, by combining multiple LoRAs - including the 8 Steps Lightning LoRA - with a new upscaling process using SeedVR2, I have finally achieved a level of quality that satisfies me.
I have also released a separate workflow dedicated solely to upscaling images using SeedVR2. You can find it here.
Update 2025-10-12: Chroma workflow added
PhotoFlow Chroma HD Flash
After Danrisi released the awesome Lenovo UltraReal LoRA, I gave Chroma another shot because the images looked super promising. And I gotta say, Chroma HD is amazing! Chroma needs about 50 steps in standard mode for awesome results, and even with an RTX 5090, that's too long of a wait for me. Fortunately, there is a Chroma Flash LoRA that achieves great results with just a few steps. With CFG 1.0, the time per step is halved again!
Update 2025-08-23: Wan 2.2/2.1 T2I workflow added
PhotoFlow Wan T2I
Wan 2.1 and Wan 2.2 are video models that can also be used to create high-resolution images. Due to the impressive quality, I decided to create a dedicated workflow for them. Wan 2.2 was released during the development of this workflow. Fortunately, Wan 2.2 is even better and can be used like the Wan 2.1 model, including compatibility with most Wan 2.1 LoRAs. Style LoRAs, in particular, work very well. SageAttention with Triton is supported but not required, and the Lightning LoRA is used to ensure fast generation of high-resolution images. Instructions on how to set up and use the workflow are included within the workflow itself.
PhotoFlow XL
This SDXL ComfyUI workflow has been developed to generate realistic photos with a focus on people.
The version 2.0 is not restricted to the use of the DMD2 Speed LoRA. The workflow works just as well without it ♥️.
The default settings and comments in the workflow are geared towards DMD. Accordingly, you only need to adjust the CFG, sampler, scheduler and steps for a non-DMD use 🙂. These values are a good start for a generation without the DMD LoRA:
KSampler
CFG: 3.5 (2.5 - 5.5)
Sampler: dpmpp_sde_gpu
Scheduler: normal
Steps: 25
FaceDetailer
CFG: 3.5 (2.5 - 5.5)
Sampler: dpmpp_sde_gpu
Scheduler: karras
Steps: 18
Recommended Checkpoints
The following checkpoints are particularly suitable:
SFW
epiCRealism XL: https://civarchive.com/models/277058?modelVersionId=1920523
HelloWorld XL: https://civarchive.com/models/43977/leosams-helloworld-xl
Juggernaut XL: https://civarchive.com/models/133005/juggernaut-xl
RealVisXL 5.0: https://civarchive.com/models/139562?modelVersionId=789646
NSFW
Jib Mix Realistic XL: https://civarchive.com/models/194768?modelVersionId=1809082
DonutsDeliveryMix: https://civarchive.com/models/605283?modelVersionId=1739105
Lustify: https://civarchive.com/models/573152?modelVersionId=1094291
Analog Madness SDXL: https://civarchive.com/models/408483/analog-madness-sdxl-realistic-model
Big Love XL2: https://civarchive.com/models/897413?modelVersionId=1464711
XXX-Ray v1.1: https://civarchive.com/models/1064836?modelVersionId=1305584
NSFW (DMD included)
GonzaLomo (Unity/Real XL DMD): https://civarchive.com/models/1513492/gonzalomo-xlfluxpony
Lustify ENDGAME DMD2: https://civarchive.com/models/573152?modelVersionId=1099200
XXX-Ray DMD2: https://civarchive.com/models/1064836?modelVersionId=1624818
Installation
ComfyUI Manager Installation
https://github.com/ltdrdata/ComfyUI-Manager?tab=readme-ov-file#comfyui-manager
Installation of Missing Nodes
Missing nodes in this workflow are highlighted in red. To install them, open the ComfyUI Manager and click on the “Install Missing Custom Nodes” button.
⚠️‼️ComfyUI-KJNodes needed‼️⚠️
It it may happen that the ComfyUI-KJNodes are not included in the list of missing nodes. Simply install the ComfyUI-KJNodes via ComfyUI Manager > Custom Nodes Manager.
After the installation, restart ComfyUI and refresh the browser page.
Further information and tips on configuration can be found in the workflow in the INFO node.
TL;DR
The workflow was developed and tested with a NVIDIA RTX 4080 (16 GB VRAM) and 96 GB RAM.
Leave a positive review, if you like the workflow 😊.
Description
Qwen-Image-2512 offers exceptional prompt comprehension and delivers outstanding output quality. However, this comes at the cost of very long generation times due to the model's size and the large number of processing steps required. A practical solution in this workflow is to use a heavily quantized FP4 version, which maintains surprisingly high quality. When combined with an 8-step Lightning LoRA, generation speed improves dramatically while image quality remains excellent.
FAQ
Comments (21)
It's funny that nunchaku doesn't work with a 14Gb model on GF4080/16 - it calls OOM in ksampler, and the standard ComfyUI model loader works fine.
Trying to play around with qwen for the first time, using the 2512 nunchacku v1 wf with the: "nunchaku_qwen_image_2512_best_quality_int4.safetensors"
Model, generation times are taking ~30 minutes for a single image, (System: 16gig vram/64ram).
Am I using too high a model for my system or am I missing something? Using sageattention generates black boxes for me even with both of the patch node active so that's disabled for these runs. .
Which GPU do you have exactly? Are your drivers up-to-date? Which ComfyUI version is installed? Do you use Windows, Linux or Mac? 30 minutes is too long.
@ai839 RTX 4070ti super, drivers up-to-date, Windows 10.
Comfyui 0.19.3, python 3.13.11 2.11.0+cu130.
I tried the Photo flow Qwen V1 wf and it was significantly faster. (~8 minutes, most of that the SeedVR2 nodes).
On the Qwen v1 wf I disabled the SeedVR2 upscale at it finished in 70 seconds.
@srfuentes99 Where exactly does the Qwen-2512 workflow slow down? Is it already in the first KSampler? Try the smallest nunchaku model. The quality is still very good. Reduce the initial image size. Watch the VRAM consumption in the Performance tab of the Windows Task Manager and try to find a configuration in the workflow where all fits in the VRAM (maybe not possible). It starts to get slow if data is swapped to RAM or even SSD/hard drive.
@ai839 problem seems to be the num_blocks_on_gpu. When it was set to 60 GPU usage was completely maxed out. Dropped the number to 30 and gpu usage never went above 30%. Finished in ~5 minutes on the initial run and significantly faster after.
@srfuentes99 Thanks for sharing the solution!
@srfuentes99 I have your identical problem! I also have 16gb VRAM (RTX 5070 TI) + 64gb DRAM DDR5.
As you suggested, I set num_blocks_on_gpu on 30. Got slightly better but still slow. Did you find any other better solutions in the meanwhile?
@giuri36 Sorry, I didnt really test past that. I decided Qwen wasnt worth the trouble and went back to zimage turbo and klein.
@srfuentes99 I understand, I'm having too many issues as well! It's a pity as Qwen Nunchaku seems to be a good model, and I like comparing its outputs with Flux and ZiT! Hope @ai839 can find us a solution!
@giuri36 You can contact me in the chat. Maybe I can help.
Can you do version without nunchaku ?? as have everyday problem with it... each update comfyui or any nodes make nunchaku not work.. and wasting halfd day trying to fix it..
I don't think I'll ever find the time to do that.
damn, that´s sad. Given that the nunchaku one is completely brroken since the last update, at least for me. I´m very thankful for all the awesome stuff you put out, so of course noone is in a position to complain, but if you ever do find the time, it would be awesome. It´s the best workflow for qwen out there imho.
@fanboytunes0789 pytorch in ComfyUI portable was updated from version 2.11 to 2.12. Sadly there is no Nunchaku support for Windows for this version. The solution is to downgrade to pytorch 2.11:
1. Replace the ComfyUi\comfy\ldm\qwen_image\model.py file with https://raw.githubusercontent.com/Comfy-Org/ComfyUI/v0.22.0/comfy/ldm/qwen_image/model.py
2. Open the Explorer and open the python_embeded folder in ComfyUI.
3. Type "cmd" in the address bar to open the terminal.
4. Execute the command: .\python.exe -m pip uninstall -y torch torchvision torchaudio
5. Execute the command: .\python.exe -m pip install torch==2.11.0+cu130 torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130
Now nunchaku can be installed again.
@ai839 Wow, thanks for taking the time to give such detailed help. I can only try it tomorrow, but it makes sense that would work. Awesome!
If anyone is interested in a simple Krea 2 workflow (embedded in the image): https://civitai.red/images/134971407
yeeee thank you , i was about to come and comment about ideogram and krea 2
An even better and faster Krea 2 Turbo PhotoFlow workflow was released.
@ai839 thats so nice