π· HD IMAGE OUTPUT β Now with ControlNet & Img2Img! (Up to 2048P)
π SkyReels and WAN β Text-to-Video Models, Stunning Still Results
Sharing my current wrapper workflows for WAN and SkyReels β both are text-to-video models, but they generate incredible still images too.
βοΈ Defaults work great, but everythingβs tweakable.
π Donβt skip the in-workflow notes β they cover setup and tips.
π‘ Running with under 16GB VRAM? Enable block swapping for smoother runs.
π Want faster speed? Drop steps to 4. The examples here used 10 steps and ran in ~25 seconds on an RTX 5090.
π₯ After lots of testing, Iβm getting better results than FLUX β faces are more realistic, and no FLUX face issues.
To Simplify things you can just swap models out instead of having two separate workflows for each model. The links to the models are in the workflows.
βWhatβs the difference between these two workflows?
Glad you asked! After plenty of testing:
π¬ SkyReels gives a slightly more cinematic and realistic look.
π WAN is a bit sharper and punchier, but not as cinematic.
That said, LoRA choice heavily affects output, so I recommend trying both workflows with your prompts to see which one fits your style best.
π§ These features currently work with the WAN wrapper:
ποΈ Text-to-Image
π§ ControlNet (DepthAnything, DW Pose, Canny, Lineart, etc.)
πΌοΈ Image-to-Image (enhance/upscale with denoise control)
β
Native support is available now β ControlNet and Image-to-Image coming soon!
β
Native GGUF support is available now β ControlNet and Image-to-Image coming soon!
π Why do these look so good? Simple: theyβre trained on video, meaning way more frames per concept β and WAN is a 14B model, much larger than FLUX. That scale and frame data really shows in the results.
Give them a spin and let me know what you think!
π Need help crafting text-to-image prompts? I made a GPT just for that: Perfect Text-to-Image Prompt Creator
Try them out and let me know what you think!