Sign In
homemodelsimagesvideospostsarticlesbountieschallengeseventsupdatesshop

[ NOTE πŸ“Œ] about working with the PonyV6 model.

9

Magnu_s's Avatar

Magnu_s

Apr 11, 2025

(Updated: a year ago)

musing
[ NOTE πŸ“Œ] about working with the PonyV6 model.

{Draft} - This article is a short note about my work with the Pony Diffusion V6 model.

πŸ¦„ Pony Diffusion V6

The starting parameters are recommended by the authors and the community to achieve the best quality of Pony Diffusion V6 images: resolution ~1024px, 25 steps, CFG ~7, sampler DPM++ 2M (or Euler a), CLIP skip 2, its own VAE models.

The standard image sizes are usually as follows:

1024Γ—1024, 896Γ—1152, 1152Γ—896, 832Γ—1216, 1216Γ—832

  1. Euler a \ CFG Scale: ? \ Steps: 25

  2. Euler a \ CFG Scale: 7-9 \ Steps: 25-30, 45 - provides a great combination of detail and creative variety of results.

  3. DPM++ 2M SDE Karras \ CFG Scale: 7-9 \ Steps: 30-50 - to unlock the potential, but allow you to get very detailed results.

In the Pony Diffusion V6 XL model, the tags score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up are used to indicate the desired quality level of the generated images. These tags were added during model training to evaluate the aesthetic quality of the images:​

  • score_9 - highest quality images.​

  • score_8_up - high quality images.​

  • score_7_up - medium quality images.​

  • score_6_up - below medium quality images.

  • score_5_up - ?

  • score_4_up - ?

They are usually arranged in this order:

Standard:

  • score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up,

Other:

  • score_9, score_8_up, score_7_up, score_6_up,

  • The first two options are typically used. I have also noticed that, when using LoRa, there are times when it is not necessary to use any of the specified tags.

Negative prompt:

  • score_6_up, score_5_up, score_4_up

    I usually set such parameters, but it's not always possible to say that they're really necessary.

  • I also noticed that any negative suggestion affects the style of LoRa.

rating_safe - safe, family friendly, unobtrusive content.

rating_explicit - nsfw content.

rating_questionable - something in between.

source_anime - create an anime style.

source_cartoon - create a cartoon style.

source_furry - creating horse and humanoid hybrids.

source_pony - generate a horse.

lineless art -

pixel art -

BREAK - is a separator word. It tells the model: This part is finished, now start something new.

Some UI (for example, ComfyUI or A1111 with custom extensions) may interpret BREAK as a label for a new semantic block, but this behavior is not built into the model itself, but is implemented at the user interface or plugin level.

In SDXL (Stable Diffusion XL) and other modern BREAK models itself does not mean anything β€” it simply turns into a normal token like any other word.

Imagine this:

You want to draw two ponies β€” Twilight Sparkle and Rainbow Dash β€” in one picture, but clearly separate, so they don’t blend into one weird pony with both wings and a horn.

You write:

score_9, score_8_up, score_7_up, twilight sparkle, smiling, BREAK rainbow dash, flying

The model will:

  • draw Twilight Sparkle first (the first part),

  • then Rainbow Dash (the second part),

  • and won’t mix them together.

When to use BREAK:

  • You have two characters and want them to be separate.

  • You want to change the scene or focus.

  • You want to give each part its own details.

When NOT to use it:

  • If everything is about one character.

  • If you want the image to feel like one single scene.

Most common quality tags:

  • masterpiece - tells the model to make it look like a work of art.

  • best_quality - Increases sharpness and overall detail.

  • highres - aims for high-resolution output.

  • ultra-detailed, finely detailed - adds lots of small, precise details.

  • 8k, 4k, photorealistic - makes it look super clear, sometimes like a real photo.

  • cinematic lighting - adds movie-style lighting.

  • dramatic lighting - adds shadows and strong light contrasts.

  • sharp focus - makes the subject super clear.

  • beautiful lighting - adds pleasant and aesthetic light effects.

  • depth of field - blurs the background, like a pro camera shot.

Negative words that can help:

  • ugly, tiling, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, extra limbs, disfigured, deformed, body out of frame, bad anatomy, username, watermark, signature, text, cut off, low contrast, underexposed, overexposed, bad art, beginner, amateur, distorted face, nudity, nsfw, low quality, worst quality, jpeg artifacts, blurry, bad eyes, ugly eyes, (mutated hands and fingers:1.4),

9