a backup of my local (experimental, opinionated) changes to Forge2 webUI

• broadly similiar package requirements to original Forge - no new Python, Numpy, etc. I'm using Torch 2.6 and CUDA 12.4 on a GTX1070.
• auto selection of VAE and text encoders per model / UI setting
• Chroma (based on https://github.com/croquelois/forgeChroma) and Chroma Radiance (based on https://github.com/maybleMyers/chromaforge)
• extended Checkpoint Merger, including UI for nArn0's embedding convertor (based on https://github.com/nArn0/sdxl-embedding-converter)
• Hypernetworks
• various and multitudinous other tweaks and modifications: UI, embedding filtering, code consolidation and tidying, dead code removal, performance improvements (for me), improved memory management
• tiling (sd1, 2, xl) (based on https://github.com/spinagon/ComfyUI-seamless-tiling)
• all embeddings everywhere all at once: SD1.5 embeddings (CLIP-L only) can be used with SDXL, SD3; SDXL embeddings can be used with SD1 (applies CLIP-L only, CLIP-G ignored), SD3
• new preprocessors for IPAdapter, including tiling, noising (for uncond) and sharpening of inputs. And multi-input.
• Latent NeuralNet upscaler by city96 (based on https://github.com/city96/SD-Latent-Upscaler) (1.25x, 1.5x and 2.0x; for sd1.5, sd2, sdxl).
• Flux2 latent NeuralNet upscaler by TensorForger (based on https://github.com/tensorforger/comfyui-flow-upscaler) (2.0x; for Flux.2, ERNIE, mugen).
• support more upscaler models via updated Spandrel - place models in models/upscaler.
• ResAdapter support: download models to models/other_modules, load via 'Additional modules' selector (as VAE, text encoder), LoRA as usual (https://huggingface.co/jiaxiangc/res-adapter)
• long CLIP
• distilled T5 models for Flux by LifuWang (see https://huggingface.co/LifuWang/DistillT5)
• lama and MAT inpainting models usable in img2img, both as processing options and as infill options
• PuLID (sdxl) (based on https://github.com/cubiq/PuLID_ComfyUI/)
• nVidia Cosmos predict2 t2i (only tested 2B model), and Anima (including canny controlnet, edit loras (use through controlnet))
• Wan 2.2 14B (based on Haoming02's implementation in ForgeNeo) and 5B, 1 frame generation only
• ELLA (https://github.com/TencentQQGYLab/ELLA), models downloaded automatically for first use, enable in Settings
• extra samplers: SA-Solver, ER-SDE, Adaptive-ODE (params in Settings), Fixed-ODE; and extra options/variants.
• assorted extra functionality in built-in extensions, controlnet preprocessors, Extras tab.
• built-in latent manipulations such as Epsilon Scaling, CFG rescale, CFG normalization (see Settings)
• dyPE high resolution generation for Flux.1 (based on https://github.com/guyyariv/DyPE), works (inconsistently) as far as I can test on my limited hardware. Also implemented for Z-Image. And a simpler alternative method for Flux.1, Z-Image, ERNIE. See Settings > dynamicPE.
• Lumina2
• Z-Image-Turbo (1024x1024, fp8 model - 15s per iteration ⇒ 2:15 for 9 step generation on old laptop with 8GB VRAM GTX1070); including Union Control. And Base.
• Mugen
• Flux.2 Klein 4B and 9B
• ERNIE
• Krea2 (and depth controlnet (use through '(Other) Control Integrated' extension; load as an Additional Module))