Anima UltraReal: Turn Anime Characters into Real Photographs, Then Accelerate to 8 Steps

UltraReal is not an image-edit pipeline with a style transfer bolted on — it is a realism-biased full fine-tune of Anima Preview 1, so the photographic result comes out of the first sampling pass. The episode pairs it with an 8-step Turbo LoRA that finally does not gut the image, plus separate face, hand and foot refinement groups. Evidence note: three workflow JSON files ship with this episode and were parsed directly. Checkpoints, LoRA filenames and strengths, sampler settings and latent sizes below come from those files. When this is useful Casting an illustrated character into a live-action-style frame without an edit model in the loop. Fast iteration at 8 steps when you need volume rather than a hero image. Repairing faces and hands that the base 1024x1024 pass gets wrong. Choosing between open anime-to-real routes: this fine-tune versus Qwen 2509, 2511 and Flux2 Klein LoRAs. How this differs from the previous build Three different jobs live in these three graphs. The UltraReal graph is the official flow: prompt, size, run, with a negative prompt that still matters because the model is not distilled. Two extra groups handle face, hand and foot repair through SAM-based segmentation. The 8-step graph is the acceleration story: one LoRA node, steps 8, CFG 1. The author's claim is that this Turbo version loses far less than the earlier 0.1 release. The comparison graph stacks three editors side by side — Flux2 Klein FP8, Qwen-Image-Edit-2509 and Qwen-Image-Edit-2511 — each with its own accelerator LoRA, so you can look at them in one run. Running it Pick the graph for the job: UltraReal for a single hero image, the 8-step graph for volume, the comparison graph to choose an editor. Write the positive prompt. For the accelerated comparison route, the author feeds a JSON-structured prompt through an LLM node so the output obeys both UltraReal and Anima conventions. Set the size. 1024x1024 in the UltraReal graph, 512x512 in the 8-step graph. For the 8-step route, load the Turbo LoRA, set steps to 8 and CFG to 1. Run the repair groups when faces, hands or feet come out broken. Compare across editors, keep the better of each pair — the comparison graph emits several results per run. Parameters, and where the numbers come from Item Value in the shipped graphs UltraReal UNET ultrarealFineTune_v10AnimaPreview1.safetensors, weight dtype default 8-step UNET waiANIMA_v10（anima3）.safetensors Comparison UNETs fluxKleinFP8_flux2KleinBase9bFp8.safetensors (fp8_e4m3fn), Qwen-Image-Edit-2509_fp8_e4m3fn.safetensors, qwen_image_edit_2511_fp8mixed.safetensors Text encoders qwen_3_06b_base.safetensors (qwen_image), qwen_3_8b_fp8mixed.safetensors (flux2), qwen_2.5_vl_7b_fp8_scaled.safetensors VAEs qwen_image_vae.safetensors, flux2-vae.safetensors UltraReal sampler KSampler, steps 50, CFG 4, er_sde / beta, denoise 1 UltraReal latent 1024x1024, batch 1 8-step LoRA anima-turbo-lora-v0.1.safetensors at 1.0 8-step sampler ClownsharKSampler_Beta: eta 0.35, linear/euler, beta57, steps 8, CFG 1, denoise 1 then 0.55 8-step latent 512x512 Comparison accelerators klein_9B_Turbo_r128.safetensors, Qwen-Image-Edit-2509-Lightning-8steps-V1.0-fp32.safetensors, Qwen-Image-Edit-2511-4steps-bf16.safetensors (all at 1.0) Comparison anime-to-real LoRAs f2k_anything2real_a_patched.safetensors, Anything2Real_Alpha.safetensors, anything2real_2601_A.safetensors (all at 1.0) Comparison samplers steps 4 / CFG 1 / euler / beta57; steps 20 / CFG 1 / euler / simple; steps 8 / CFG 1 / euler / beta57 The 8-step graph deliberately runs at 512x512 — it is a draft-tier graph, not a final-quality one. The step count and CFG in the narration match the file exactly: 8 steps and CFG 1, with a 0.35 eta on the special sampler. Honest limits The face problem is a training-data problem: the author says the fine-tune saw too much non-Asian facial data, so Asian faces come out weaker. 8 steps costs prompt adherence and fine detail; the author describes results as looking unfinished but style-consistent. Anima Preview 1 still produces breakdowns, which is why the official flow carries repair groups at all. The author's pick for anime-to-real editing is the Qwen-Image-Edit-2511 LoRA; Klein held identity poorly, and 2509 responded weakly even with its LoRA loaded. All quality judgements here are the author's visual comparisons, not scored evaluations. Episode resources RunningHub workflows Anima UltraReal realism + face/hand/foot repair — https://www.runninghub.ai/post/2052302245529694209/?inviteCode=rh-v1111 Anima 8-step fast generation — https://www.runninghub.ai/post/2052343882582634497/?inviteCode=rh-v1111 Anime-to-real multi-model comparison — https://www.runninghub.ai/post/2052281071861026818/?inviteCode=rh-v1111 Video tutorial (YouTube): https://www.youtube.com/watch?v=23YfGpuLgLI Chinese original (Bilibili): https://www.bilibili.com/video/BV1d8dPBCErN