FreeManAi

For beginners, see their favorite diagrams and styles want to practice a model to use, then use LORA training!First of all, determine what type of LoRA you need to train, and see what you like: see a beautiful picture, what you like in the end, and want to practice in the direction you like. Roughly divided into: characters, painting style/style, concept, clothing, objects/specific elements (light, dark, fire and ice, colors, etc.).The second is to gather material. According to the elements you like to collect, against the above favorite categories to collect materials, not the more the better, the general recommendation is about 20 pieces. Material requirements: HD, different angles, different backgrounds, different poses, different clothing, clear and unobstructed pictures, the key is HD!! So the real thing please find this object to take pictures of their own best.The third step is to process the material, of course, mainly by Photoshop. If you don't know Photoshop, cut it! Clean up the things you don't like and keep the HD parts you need. Blurring, text, etc. are not recommended.The fourth step to generate/optimize the label (optional), the website generally with the tag function, so no more say, with the tool can (it is recommended to keep a txt file, easy to check the LORA effect).Step 5 Trigger word: Choose a personalized English word (to be unknown to the translator!!) , add the trigger word to the tag of each graph.Step 6 Select the large model used by LoRA for training: currently there are SD1, SDXL, PONY, FLUX, SD3.5 (the price of computing power here is from low to high), other such as (mixing elements, graphs, etc.), optional middle sample graph (magic words, please refer to the label of a previous figure)The seventh step begins training, refining on the website is the consumption of computing power, after the deduction of high and low, it is waiting for training, so that after training to see the progress bar, look at the middle of the sample diagram... This will take a long time, just wait slowly. (Usually throw it in and wait for one night)Step 8, of course, harvest a bunch of refining results, start testing the model!From the trained LoRA model queue, find the last file to download, of course (the middle one is also OK), the local SD/ComfyUI into the folder (SD-webu-aki /models/lora), After refreshing, you can see the model you trained in SD,FLUX/SD3.5 must be uploaded and published first, you can first send a private, regeneration map.According to the TXT label of their own figure, the good alchemy bottom mold, through different weights to check the LORA effect, is not very similar? Daaaaaa!----------------Experience: Comparison of SD1, fuzzy graphics (corresponding to a good bottom model)SDXL: Find the bottom mold, basically similar, with general clarityPONY: NSFW weapon, it is relatively fuzzy, the skin is 3D silicone, can not see the hair, pores of the kindFLUX: HD but not like it, unless using F1.dev-F16, it looks like it. Too much video memory, the site is only F8, so unlike, often output 2D graph, tired ~SD3.5: The style is similar, the clarity is also OK, not NSFW