我在使用mossimage炼制lora过程中发现，无论是加高步数，调整位移，调整rank，调整训练过程加入modulation+embed，均无法改正生成图片结果被强行掰回平涂。最终定位到问题很可能出在使用的vae，flux2本身上。本工作流采取了两种方案，经本人测试均可行，一种是对潜空间进行放大增强对比度，另一种是直接舍弃flux2的vae改用英伟达pid，此外还有简单的种子锁定以及选择upscale模型如1x，4x的功能，以及可选的seedvr2超分。
The problem

While training a MOSS-Image LoRA, flat-shading regression turned out to be stubborn. Higher step counts didn't fix it. Shift adjustments didn't fix it. Rank changes didn't fix it. Injecting modulation and embedding during training didn't fix it either—the output kept collapsing back toward cel-shaded, washed-out midtones.

The root cause pointed at the VAE side, not the LoRA or the sampler: FLUX.2's VAE, when fed latents from MOSS-Image after third-party style LoRA injection, frequently produces compressed variance in the latent, which the decoder then resolves into flat color and dull skin. NVIDIA PiD offered a way around it. PiD is a pixel-diffusion decoder that reformulates latent decoding as conditional pixel diffusion—it denoises directly in high-resolution pixel space and unifies decoding and upscaling into one module, producing 4× output without going through a conventional VAE decoder first. In practice, that means the decoder is no longer reconstructing midtone-heavy latents into flat shading; it is synthesizing high-frequency detail in pixel space from the latent's structure.

The fix (two paths, both tested)

This workflow includes both viable escape routes, selectable from the global control panel:

1. Latent contrast boost (keep FLUX.2 VAE): A LatentMultiply node before VAE decode. Multiplier 1.03–1.08 (default 1.05) restores enough latent variance that the FLUX.2 VAE stops collapsing into flat color. If you stay on this path and still see slight washing-out, the workflow also includes a latent pre-upscale of 1.1× before VAE entry—this was the second half of the breakthrough, and it is already wired in.

2. NVIDIA PiD 4× decode (bypass FLUX.2 VAE entirely): Feed the sampled latent into the PiD path and decode directly to 4× resolution. No FLUX.2 VAE in the loop, no flat-shading regression. PiD runs in 4 distilled steps on the fast checkpoint variant, or 12–15 steps on the original for maximum stability.

What you should actually use

• LoRA strength: 1.1–1.2. At 1.0 the effect is not strong enough on this model family. Start at 1.1; push toward 1.2 only if your prompt already has strong lighting descriptors.

• Negative prompt (recommended): flat color, 2d, cel shading, bad lighting, dull skin, matte, muted colors, low contrast, simple shading

• Turbo LoRA compatibility is poor. Turbo LoRAs will pull the output back toward flat shading regardless of which decode path you choose. Avoid stacking them on this workflow if you want the non-flat look preserved.

• Shift and 22-layer skip percentage both influence flatness. If the result still reads slightly flat after the VAE fix, nudge shift before changing anything else. The workflow exposes these in the 02 sampling group.

What else is in the workflow

• Seed lock for reproducible sampling across decode paths.

• Upscale model selector (1× skin-texture repair or 4× fast upscale), independent of the PiD route.

• Optional SeedVR2 controlled-ratio super-resolution for a final 1.5×–2.5× pass on top of either decode path. The ratio is entered as a multiplier; the target short-edge pixel value is computed automatically as an even number, so you are no longer guessing what resolution = 1080 actually means for an 832×1216 source.

• Group-level bypass control panel (rgthree Fast Groups Bypasser): turn entire sections on or off without triggering model checks on disabled groups.