moody-gal-weiwei-krea2-zimage/info.txt

═══════════════════════════════════════════════════════════════════════════════
  MOODY GAL - WEIWEI (KREA2 / ZIMAGE)
═══════════════════════════════════════════════════════════════════════════════

───────────────────────────────────────────────────────────────────────────────
  IDENTITY
───────────────────────────────────────────────────────────────────────────────
  Base model        Krea 2
  Generation mode   not declared by the author
  Checkpoint        not determined — verify before use
  Category          Character / subject
  Author            catlover1937
  Source            https://civitai.com/models/2669955
  Provenance        CivitAI · model 2669955 · version 3218400

───────────────────────────────────────────────────────────────────────────────
  WEIGHTS
───────────────────────────────────────────────────────────────────────────────
  File              Weiwei_epoch_9.safetensors
  Version           Krea2
  Size              327.0 MB
  Format            SafeTensor / unknown precision
  SHA256            bfde4c4330e85860fdcaab61fcc5b5523c8fba410dd1a88ca6fbbaca06b55c10
  Published         2026-08-14

───────────────────────────────────────────────────────────────────────────────
  TRIGGER WORDS
───────────────────────────────────────────────────────────────────────────────
  › "Weiwei"

───────────────────────────────────────────────────────────────────────────────
  RECOMMENDED SETTINGS
───────────────────────────────────────────────────────────────────────────────
  The author did not publish specific settings.
  No author settings were published. Community baseline: steps Turbo: 8
  (official README and HF card CLI example; the ComfyUI template ships 8).
  Raw: 52 (official README). The krea-2 CLI's generic default is 28.
  Community Turbo LoRAs commonly suggest 8-12 (e.g. the Identity Edit card
  recommends Turbo at 8-12, or Raw at 20 for removals). A community 4-step
  LoRA on Turbo makes 4 valid; its author calls 2 steps an out-of-spec
  preview mode., cfg Turbo: CFG disabled - guidance_scale 0.0 in diffusers
  and --cfg 0.0 in the krea-2 CLI, which is cfg 1.0 in ComfyUI's KSampler
  (the official template ships cfg 1). These are the same thing; do not set
  ComfyUI cfg to 0. Raw: CFG 3.5 (official README; CLI generic default 4.5).
  Negative prompting on Turbo is effectively unused - the official templates
  route ConditioningZeroOut into the negative slot., sampler ComfyUI:
  KSampler with sampler euler, scheduler simple, denoise 1.0 (verified in
  image_krea2_turbo_t2i.json: [seed, 'randomize', 8, 1, 'euler', 'simple',
  1]). Timestep shift mu = 1.15 for Turbo; ComfyUI bakes this in via
  Krea2.sampling_settings = {"multiplier": 1.0, "shift": 1.15}, so the plain
  t2i template needs no ModelSampling node, while the style-reference
  template adds ModelSamplingFlux at max_shift 1.15 / base_shift 0.5. The CLI
  exposes --mu 1.15 for Turbo; Raw uses a resolution-derived mu interpolated
  between --y1 0.5 and --y2 1.15 (ai-toolkit mirrors this as base_shift 0.5 /
  max_shift 1.15 over image_seq_len 256-6400, exponential time shift).,
  resolution 1024x1024 default. Turbo: 1K-2K (the official CLI example
  generates 2048x2048; ComfyUI's ResolutionSelector defaults to 1 megapixel,
  set 2.0 for 2K). Raw: "trained to generate upto 1k resolution".
  Width/height are padded up to a multiple of 16., lora_strength 1.0 for
  Krea's nine official style LoRAs (docs.comfy.org lists Recommended Strength
  1.0 for every one), though the shipped t2i template's LoraLoaderModelOnly
  is set to 0.8. The style-reference LoRA is used at 1.0 in its template. The
  4-step distill LoRA documents 1.0 as its trained point, 1.0-1.5 as coherent
  extrapolation and above ~1.5 as speckle artefacting. Community subject
  LoRAs are typically recommended around 0.8-1.2 and dropped toward 0.6-0.9
  when several are stacked..

───────────────────────────────────────────────────────────────────────────────
  ADOPTION & POPULARITY
───────────────────────────────────────────────────────────────────────────────
  Downloads               1,819
  Rating                  100% positive (95/95)
  Comments                0
  Published               22 days ago
  Download rate           30.9/day since release

  Adoption rank           #123 of 200 in this collection  (Modest)
  Momentum rank           #134 of 200 by download rate  (Long tail · fresh release)

  Adoption tier is the percentile of total downloads within this collection;
  momentum is the percentile of downloads-per-day since release. Momentum is
  ranked rather than measured against a fixed rate, because this base model
  is itself new and every LoRA looks fast on an absolute scale. Both numbers
  are CivitAI's, read on the scrape date at the foot of this file.

───────────────────────────────────────────────────────────────────────────────
  CHAINING / STACKING
───────────────────────────────────────────────────────────────────────────────
  Stacks with others      not stated

  The author gave no chaining guidance. The rules below are the
  general Krea 2 ones, not anything specific to this LoRA.

  General Krea 2 rules that apply here (see REFERENCE.txt):

  • The most important rule is not about chaining but about which
    checkpoint you chain onto: the official README says in bold "TRAIN on
    Raw and RUN on Turbo", and "We highly recommend using RAW for training
    LoRAs and applying them on Turbo for inference." LoRAs trained on Raw
    are stated to work well on Turbo.
  • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2
    LoRAs are DiT-only, so there is no CLIP side to chain and the text-
    encoder branch stays untouched however many LoRAs you add.
  • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as
    routine - style + character, style + slider, realism + character.
    Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one
    writes "I recently stacked over 20 of my own LoRAs on a single
    generation and the output still came out great."
  • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge
    on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is
    loaded. The CivitAI 12GB training guide also notes rank affects
    composability: "rank 8 produced a lighter style influence that was
    easier to stack with other LoRAs, rank 16 provided a useful balance
    between concept strength and composability, and rank 32 was better
    suited to a precise object or concept intended to take priority over
    other stacked LoRAs."
  • Accelerator LoRAs are position-sensitive in intent, not in node order:
    a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a
    sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to-
    Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is
    explicitly recommended by the 4-step LoRA's author.
  • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI
    (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card
    says "Keep mu = 1.15. The training targets are anchored to that grid; a
    different shift moves the [sampler off them]"; the same reasoning
    applies to any Turbo-trained adapter in the chain.
  • Reference/edit LoRAs are designed to be stacked under ordinary subject
    and style LoRAs - the Identity Edit card states "Composes with your
    LoRAs: character/body/style LoRAs stack on top and steer the prior" and
    recommends a stacked subject LoRA when unusual subjects drift toward
    the base prior.
  • Krea 2 LoRAs cannot be chained with LoRAs for any other base model.
    Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or
    "[Flux | ZIB | Krea2]" - those are separate files for separate
    architectures shipped under one model page, not a combined adapter.
  • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI
    ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per-
    section weights for first., tmlp., txtmlp., tproj.,
    txtfusion.projector, txtfusion.layerwise_blocks.0-1,
    txtfusion.refiner_blocks.0-1, blocks.0-27 and last.
  • The author never states which weight family this targets. Silent
    degradation, never a hard error, between Raw and Turbo. The two share
    one architecture - ai-toolkit's config is commented "The reference
    'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)"
    - so every LoRA tensor shape matches on both and ComfyUI/diffuser

───────────────────────────────────────────────────────────────────────────────
  PROS
───────────────────────────────────────────────────────────────────────────────
  + Well established — 1,819 downloads. [source: CivitAI stats]
  + Positively received — 95 thumbs-up with no down-votes. [source: CivitAI
    stats]
  + Explicit trigger word(s) — Weiwei — so the effect can be turned on and
    off from the prompt. [source: model metadata]
  + Compact at 327 MB — cheap to stack with other LoRAs. [source: file
    size]

───────────────────────────────────────────────────────────────────────────────
  CONS & CAVEATS
───────────────────────────────────────────────────────────────────────────────
  - Author never states which generation mode it was trained for; test
    against your own workflow before relying on it. [source: description]

───────────────────────────────────────────────────────────────────────────────
  COMPATIBILITY WARNING
───────────────────────────────────────────────────────────────────────────────
  Silent degradation, never a hard error, between Raw and Turbo. The two
  share one architecture - ai-toolkit's config is commented "The reference
  'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so
  every LoRA tensor shape matches on both and ComfyUI/diffusers will load an
  adapter against either without complaint. The lvladikov 4-step LoRA card
  states this for the reverse direction: "Every layer it targets also exists
  in Krea 2 Raw, so it will load there without complaint - but that is a side
  effect of the shared architecture, not a supported mode", and reports
  mixed, subject-dependent results with malformed subjects (duplicated heads,
  fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before
  output is coherent. The intended direction (train on Raw, run on Turbo) is
  officially blessed. The more common practical failure is a settings
  mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at
  Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output
  with no warning. Loading a Krea 2 LoRA against a different model family
  entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs
  "lora key not loaded: {key}" per key and generates as if no LoRA were
  applied.

───────────────────────────────────────────────────────────────────────────────
  AUTHOR'S DESCRIPTION
───────────────────────────────────────────────────────────────────────────────
  I just made a new Telegram Group and Discord group drop in for chats,
  questions, feedbacks or just to share your work. 欢迎加入 纸飞机 和 Discord 群组。
  喜欢我的作品吗？🎨 我会持续免费分享更多优质内容✨，欢迎支持一下，请我喝杯咖啡吧！💖 Like my work? 🎨 I share all my
  work for free ✨, so support me by buying me a coffee 💖!
  Ko-fi: https://www.ko-fi.com/catlover1937

  RH体验地址（邀请码Invite Code: cptpovgo）： https://www.runninghub.ai/user-
  center/1993809672580788226?inviteCode=cptpovgo
  Please do not redistribute without consent. 如非作者同意禁止搬运。
  This model is sponsored by @bip0larmain . Because he really likes weiwei
  :D. 本模型由 @bip0larmain 赞助。

  2026-08-11

  Krea2 Version Released.
  Krea2 版本放出。

  Sorry i was testing out blue buzz purchase, and now i can't cancel it, so
  worst case you wait for 3 days 😒 I am just trying to see how the blue buzz
  purchase work, just for fun, technically you can pay with blue buzz now,
  anyway. worse case you just wait 3 days :P.

  我只是純粹是為了好玩想看看藍色BUZZ幣購買Lora, 結果現在關不上了。理論上現在就可以用藍色BUZZ幣付款了。最壞的情況也就是等三天而已 :P。

  🌸 Release Note | Weiwei Character LoRA 🌸
  🇬🇧 English
  Hey everyone 👋 It's time for another character release ✨
  This time I wanted to create a character that feels a little more relatable
  and grounded in everyday life. Someone who might seem ordinary at first
  glance, but has a lot of charm hidden beneath the surface 💖
  Her name is Weiwei 🐾 She is a 27-year-old Chinese editor living in the
  bustling city of Shanghai 🌆
  Working long hours in publishing and media, Weiwei spends most of her days
  proofreading manuscripts, editing articles, and meeting impossible
  deadlines 📚☕
  Her life is stable and comfortable, but sometimes a little too routine.
  While many of her friends have already settled down, Weiwei remains single
  and often jokes that her tiny Pomeranian is her only family 🐶💕
  Outside of work, she enjoys:
  - 📖 Spending entire weekends reading novels at cozy cafés

  - ☕ Exploring hidden coffee shops around Shanghai

  - 🐶 Taking her fluffy Pomeranian on evening walks

  - 🎧 Listening to podcasts and audiobooks during her commute

  - ✍️ Secretly writing romance stories she never shows anyone

  - 🍰 Hunting for new dessert shops and bakeries

  - 📸 Photographing old streets and city corners

  - 🌙 Watching dramas late at night while wrapped in a blanket

  - 🛍️ Browsing bookstores even when she doesn't need more books

  - 🌸 Taking short solo trips whenever she feels overwhelmed

  Although she often appears calm and independent, Weiwei is actually quite
  sentimental and romantic at heart ✨
  She believes in meaningful connections, handwritten letters, and the idea
  that the right person will eventually appear when the timing is right 💌
  Her friends often describe her as elegant, reliable, and easy to talk to.
  The kind of person who always remembers your birthday and somehow knows
  exactly what to say when you're having a bad day 🌷
  Hope you all enjoy her!
  Like my previous character LoRAs, Weiwei is a completely fictional
  character and is not based on any real person ✨
  ————————————————
  🇨🇳 中文
  大家好 👋 又到了新角色发布的时间啦 ✨
  这次我想做一个更加贴近现实生活的角色。她或许没有明星般耀眼的光环，却有一种藏在日常中的独特魅力 💖
  她的名字叫 Weiwei 🐾 是一位生活在魔都上海的 27 岁女编辑 🌆
  作为出版与媒体行业的从业者，Weiwei 每天的工作几乎都离不开校稿、编辑文章以及赶各种截稿日期 📚☕
  她的生活稳定而平淡，但有时候也会觉得有些单调。身边许多朋友都已经恋爱甚至结婚了，而她至今仍然单身，经常开玩笑说自己家的小博美犬就是她唯一的家人 🐶💕
  工作之外，她喜欢：
  - 📖 在咖啡馆里看一整天小说

  - ☕ 探索上海各种隐藏咖啡店

  - 🐶 带着自己的小博美散步

  - 🎧 通勤时听播客和有声书

  - ✍️ 偷偷写一些从未发表过的恋爱小说

  - 🍰 打卡甜品店和面包店

  - 📸 记录上海老街和城市角落

  - 🌙 深夜裹着毯子追剧

  - 🛍️ 逛书店，即使家里的书已经堆不下了

  - 🌸 心情不好时来一场说走就走的短途旅行

  虽然表面上看起来独立又理性，但 Weiwei 的内心其实非常感性，也依然相信爱情 ✨
  她喜欢那些真诚而细腻的感情，也相信总有一天会遇见那个对的人 💌
  朋友们都觉得她温柔、可靠，而且特别善于倾听。无论什么时候找她聊天，她总能说出让人感到安心的话 🌷
  希望大家会喜欢她！
  和之前一样，Weiwei 也是完全虚构的原创角色，并非基于任何真实人物制作 ✨

───────────────────────────────────────────────────────────────────────────────
  EXAMPLE MEDIA
───────────────────────────────────────────────────────────────────────────────
  6 preview(s) in ./example_images/ — community results for
  this LoRA. Videos keep a matching _poster.jpg still frame.

  No prompts are recorded below: CivitAI's public API flags these posts as
  having generation metadata but does not return it, so the prompts cannot be
  scraped. Open the model page to read them.

  01_image.jpeg  (image, 2176x3840)
  02_image.jpeg  (image, 2368x3520)
  03_image.jpeg  (image, 2176x3840)
  04_image.jpeg  (image, 2176x3840)
  05_image.jpeg  (image, 2176x3840)
  06_image.jpeg  (image, 2176x3840)

───────────────────────────────────────────────────────────────────────────────
  HOW TO USE (COMFYUI)
───────────────────────────────────────────────────────────────────────────────
  1. Update ComfyUI to at least v0.26.0 (released 2026-06-23), which added
  local Krea2 support. For reference-image / style-reference / edit LoRAs you
  need at least v0.29.0 (2026-07-29), which added regular and timestep-zero
  reference images for Krea 2. ModelMergeKrea2 arrived in v0.27.0
  (2026-06-30).

  2. Put the base stack in place (files from huggingface.co/Comfy-Org/Krea-2,
  which is not gated):
     - ComfyUI/models/diffusion_models/krea2_turbo_fp8_scaled.safetensors
  (documented recommended default; the repo also holds krea2_turbo_bf16,
  _int8_convrot, _mxfp8, _nvfp4 and krea2_raw_bf16 / _fp8_scaled /
  _int8_convrot)
     - ComfyUI/models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors (bf16
  also available)
     - ComfyUI/models/vae/qwen_image_vae.safetensors

  3. Put the LoRA .safetensors in ComfyUI/models/loras/ - flat, no per-model
  subfolder required.

  4. Build or load the graph. The official image_krea2_turbo_t2i.json
  template is: UNETLoader (krea2_turbo_fp8_scaled.safetensors) ->
  LoraLoaderModelOnly -> KSampler; CLIPLoader
  (qwen3vl_4b_fp8_scaled.safetensors, type "krea2") -> text encode ->
  KSampler positive, with ConditioningZeroOut into negative; VAELoader
  (qwen_image_vae.safetensors) -> VAEDecode -> SaveImage; EmptyLatentImage
  1024x1024 driven by ResolutionSelector. An int8 sibling template
  (image_krea2_turbo_t2i_int8.json) also ships.

  5. Load the LoRA with LoraLoaderModelOnly, not LoraLoader - both official
  templates do. Krea 2 LoRAs patch the DiT only; note that the DiT's own
  text_fusion / txtfusion blocks are a normal LoRA target (the official style
  LoRAs patch them), but the Qwen3-VL text encoder itself is not. Set
  strength 1.0 for Krea's official style LoRAs (the shipped template uses
  0.8) and add the LoRA's trigger word to the prompt.

  6. Sampler settings: steps 8, cfg 1, sampler euler, scheduler simple,
  denoise 1.0. Do not add a ModelSampling node for plain t2i - ComfyUI
  already applies shift 1.15 from Krea2.sampling_settings.

  7. For style reference, use the dedicated template: UNETLoader
  krea2_turbo_int8_convrot.safetensors, LoraLoaderModelOnly
  krea2_style_reference.safetensors at 1.0, ModelSamplingFlux (1.15, 0.5),
  TextEncodeQwenImageEditPlus fed by LoadImage,
  FluxKontextMultiReferenceLatentMethod set to index_timestep_zero, and
  SamplerCustomAdvanced with KSamplerSelect euler + BasicScheduler simple/8 +
  CFGGuider cfg 1.

  8. For instruction editing (Krea 2 Identity Edit), stock nodes are not
  enough: install the ComfyUI-Krea2Edit node pack, which ships its own
  workflows, and run Turbo at 8-12 steps / cfg 1 (or Raw at 20 steps / cfg 3
  for removals) with LoRA strength 1.0.

  9. Low-VRAM alternative: install city96's ComfyUI-GGUF and use a community
  GGUF build (realrebelai/KREA-2_GGUFs, quantized from Turbo;
  molbal/krea2-gguf, quantized from Raw) with the Unet Loader (GGUF) node in
  place of UNETLoader. There is no official GGUF release. LoRA loading is
  unchanged.

───────────────────────────────────────────────────────────────────────────────
  Scraped by lorakit on 2026-09-05 03:59 UTC.
  Stats and description are the author's; pros/cons are derived from the
  evidence tagged beside each line.
───────────────────────────────────────────────────────────────────────────────