Samplers make a huge difference regardless on most models, but for text generation, set sampler to dpmpp_sde and scheduler to ddim_uniform.
compare the attached picture (with the aforementioned sampler) and this very first attempt I got with this very default comfyUI prompt (and I think euler a sampler and ‘simple’ scheduler):
That is a lot of text to get right… oh, the attached pic is also the Q8 gguf while the other one was made on the full bf16.
Using AI to generate the title text seems like way more effort than just doing the text overlay yourself. Or have the AI generate the text and a program write the overlay if you want to be fully automated.
You really want to focus on the text of the items in the image, like the text on the bike. And I’m not seeing a significant improvement there.
I’m not sure if this is how the model thinks about it; this was the prompt given (default that came with the workflow):
A magazine cover photography of a smiling energetic 16-year-old Japanese girl with layered short hair, pushing a vintage bicycle in front of a retro mint green vending machine. Cheerful expression, lively posture, summer vibe. She wears a white T-shirt and denim overalls. Green grapes and a water cup in the bike basket. Background of messy telephone poles and nostalgic Japanese shop signs. Side sunlight creating a golden halo on her hair. Fujifilm Pro 400H style, grainy film texture, low saturation, slightly overexposed, cinematic composition, unique camera angle. Fashion editorial style, 8K resolution.
Magazine cover layout with visible text:
Large title “SUMMER” at the top.
Small cover text: “Youth & Freedom”, “Tokyo Street Issue”, “Vol. 24 | August 2025”.
Barcode at the bottom corner.
As text on bike was not specified, the model filled in the gaps itself. We also note that it laid the text out itself, respecting margins and a grid consistent with a magazine cover.
In a real workflow you’d already have an indesign file of your cover where you just change the text and picture, but for a sample prompts it showcases the model’s capability (esp. as people want text generation more and more).
If I remember tomorrow I’ll spin up Comfy and see if we can get text on the side of the bike, or even detail in the vending machines.
