v0.0.0 (Initial Release)
The purpose of this LoRA is to break the AI's deeply hardcoded composition bias of placing a "girl in the center of the frame." To achieve this, it was trained on 60 images where the subject is intentionally positioned away from the center. As a result, the output strongly mimics the compositions found under the fixed-point camera.
How to Use & Tips: The trigger word fpcview alone is not enough to force the effect. To get the best results, please assist the composition by combining it with traditional perspective/positional tags, such as:
from above,wide shot,dutch anglehead out of frame,straight-onOr descriptive natural language like:
"The character stands near the edge of the frame while a large empty space remains visible."
Known Issues (Composition Conflicts): Because this LoRA strongly conflicts with the AI's default composition rules, you may encounter the following issues:
Resolution & Aspect Ratio: More than 50 of the training images are in horizontal (landscape) format. Therefore, this LoRA may not function properly if you attempt to generate images in vertical (portrait) or square aspect ratios.
The character count might unintentionally increase.
Instead of generating
1girl, 1boy, the two characters may "melt" (fuse) together, rendering only a single bizarre person.(During my own testing, this conflict once generated an image of a bald middle-aged man lying limp on a bed with an ecstatic expression. For the record, I am straight.)
Description
Initial.
dim4/alpha8, unet 1e-04(adamw8bit/cosine), warmup 200, 60images / 600 steps
FAQ
Comments (7)
More of this are needed! what a great concept.
It’s an excellent model; the new angles for the action shots break the mold of always keeping the subject centered. It works better horizontally—at least for me, it performed very well that way compared to vertical.
Got any tips? Cause I cannot seem to get this lora to do anything. From above, wide shot, distant view, full sentence descriptors or what I want, doesn't seem to make a difference, the view hardly changes. I really want to use this but I'm not having any luck. You using base Anima checkpoint?
@tymnwp
Q&A
Q1. Have you tried generating in a landscape aspect ratio?
This LoRA is much less effective at portrait resolutions.
Q2. Are you using Base v1 + Turbo v2 at 12 steps (or fewer)?
That combination tends to make compositions much more rigid. I recommend either switching to the official settings (er_sde sampler, simple scheduler, 30 steps, CFG 5), or—although it's still experimental—using the official Anima Turbo LoRA v0.2 with the Turbo model.
For reference, every image in the showcase was generated with Anima-Turbo-v1 + official Anima Turbo LoRA v0.2 at 12 steps.
Q3. Try simplifying your prompt as much as possible.
Start with only the essentials, then gradually add tags until you find the point where the LoRA stops having a noticeable effect.
Q4. Every showcase .webp contains the full ComfyUI workflow.
You can drag and drop any showcase image into ComfyUI and simply click Queue Prompt to reproduce it. If there's a composition you're interested in, I recommend loading the workflow in your own ComfyUI setup and breaking it down to see exactly how it's working.
@wraith_executioner_year_2
A1: Yes
A2: Base, Aesthetic, and a couple other checkpoints, both with and without Turbo V2, all with er_sde/simple (and a few other schedule/sample for testing)
A3: I've done some pretty basic prompts for testing, mostly just 1 subject and then positioning tags.
A4: I'll take a look at those. I use Forge Neo, I just don't like ComfyUI's UI, but I do have it installed so I'll give that a shot for testing.
I appreciate the response, and I'm sure it's user error somewhere, so I'll keep toying with it.
Absolute wonderful job!
Absolutely brilliant! Eagerly awaiting upgrades and more LoRAs in the same vein.