Wan 2.1 is increasingly becoming a must-have in T2I, hence my interest in this model.
Lora for Wan 2.1 T2V 14b-Q4-K-M model to use for Text to Image.
Wan 2.1 Text to Image Workflow (or use my WF included in my PNG files)
Install the new sampler schedulers with RES4LYF
Showcase images were created with ComfyUI and are in PNG, include metadata & workflow.
Description
Model trained on civitai with 100 images.
Recommended resources : Wan 2.1 T2V 14b-Q4-K-M
Recommended settings : res_2s / bong_tangent/ 10 steps / CFG : 1.0
Weighting : 1.0
FAQ
Comments (16)
Amazing. It's anatomically superior to flux. I'm not worried about the lack of an aesthetic of wan as there's adel ai :)
Yes, I absolutely agree with you. We're going to get nostalgic for XL's 6 fingers, Flux's goat feet, or SD3's alien anatomy.
Not only superior to flux, but also 25 times faster than Flux Pro.
It could easily become the next Illustrious/NoobAI with a lot more facial and anatomical variety as well. WAN is kicking ass lately.
This looks fantastic, but not sure I quite follow. This is intended to use Wan T2V for image generation, is that right? I was thinking this may be related to discussion I saw recently about using Wan basically for T2I by setting length to 1. If I'm right about that, what the LoadImageExtended node for?
Indeed, this model was trained only on images, so it's more intended for T2I or I2I. But I've seen it used in T2V and V2V by combining it with other more video-oriented LoRAs, and it seems to be working well.
The LoadImageExtended node is used to do I2T2I (with Florence2) and generate an oriented latent. By applying an oriented latent instead of an empty latent, we obtain much better output.
Ahh, thanks so much for the quick response. So for example your shower girl pic (with one hand raised and the other on the bar) started out as a similar image? I'm familiar with Florence for caption extraction from image, but still not quite sure how to optimize a source image in this context or what practical purpose it serves. (You say oriented latent is better, but I'm not sure of the practical implication)
@logicflint The oriented latent is absolutely huge. I almost never use simple T2I with an empty latent anymore (or bend it with my oriented latent to bring more variation). All my prompts are stored as images; it's more intuitive for searches.
You can manually modify your latent image by adding, for example, a yellow circle to simulate a sun, and you'll have the sun where you wanted it to be in your rendering.
It's more interesting to work on the latent than on the prompt.
Thanks for that explanation. Showing my ignorance here, but how is this overall functionality significantly different from Flux Redux?
It's an interesting story, so I leave a comment. Is oriented latent the concept of delivering a rough sketch or layout to an image through Florence2?
@Adel_AI Hey, checking with you re question from two days back. Is this (using oriented latent and Wan for I2T2I) effectively like a super-Redux, and if so in what ways is it expected to be better? I'm not challenging anything about it, just trying to understand. Thanks again!
@logicflint To truly grasp these workflows, you’ll need to experiment yourself. Explaining it all in theory would be long and abstract. Still, here’s a brief overview of the different approaches:
Classical method: This involves injecting random noise into an empty latent. Paired with a seed, this generates an output largely determined by the initial noise—so it’s a bit of a gamble.
Noise injection: There are several variations of this method, all of which improve rendering quality significantly.
Unsampling approach: Here, noise is extracted from an existing image using a sampler. This method starts to steer the noise in a deliberate direction.
Oriented noise approach: Latents are generated directly from an image—real photo, sketch, doodle, or anything visual—chosen based on the desired outcome. Lighting, colors, and even compositional layout are already heavily influenced, even when denoising is set to 1.0.
To further shape the output, you can fine-tune the denoising level. It’s best to keep it above 0.85—going lower leans toward image-to-image (I2I) generation, leaving the model less creative freedom.
If the base image used to generate the latent is interpretable, it can be processed by an image descriptor to extract a caption. That caption—whether raw or modified—can then be used as a prompt. That’s the essence of Image-to-Text-to-Image (I2T2I).
This last approach unlocks powerful control and customization options.
@Adel_AI All that make sense, thanks much. Still, I'm not following how this is fundamentally different or more powerful than Flux Redux for I2I (where unless I misunderstand the incoming latent is considered and an "under the hood" I2T(2I) is occurring, so much more powerful and with more creative knobs than typical Flux I2I).
To put a fine point on it, is the advantage here using Wan that there isn't the anatomy censorship of vanilla Flux, so it's basically a Wan based uncensored Redux? Or some other advantage I'm not quite grokking? Not asking for the theory, understood it's not for you to explain. 😁👍
"I'm GLUM... dabba dee... dabba doo..." played on the jukebox behind the girl, very slowly as if time was running down. "I'm GLUM... dabba dee" it began again. I reached for the .357 in my pocket and it seemed to take almost twenty minutes for my hand to reach the grip in my underarm holster. I drew, seemingly taking another ten minutes. The jukebox played slower, ever slower; "I'm GLUM". The girl's sad eyes locked with mine and I could sense the pleading in her gaze. It was in my hand now and I watched the moving barrel ease itself through the air like a six-hundred thousand tonne cruiseliner turning in the ocean. Another twenty minutes appeared to tick by and I was aiming at the jukebox. Five minutes later, the trigger finally depressed and I watched the round ease its lazy way from the muzzle toward the jukebox. Time was slowing, but the bullet kept moving. It ruffled the air. It was followed by pretty flame in the air. Patience was required. The bullet struck the jukebox and I saw it slowly penetrate the glass, the steel, the record, the.... All holy jesus of all the moments I'd ever experienced there folded over into one single moment and we were free. She fell forward into my arms. I grabbed her. Normality had resumed. Time passed rapidly now, or so it seemed after the slow version. Normality returned. She wept in my arms. We were free. She smiled.
Excellent WAN Work. Thank You.
RES4LYF is the best damn sampler ever. Without question.
Wan Female Nude Art.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.