# DnD-K2T-T2I v1.8 — "Dust and Dreams"

Krea 2 Turbo text-to-image workflow for ComfyUI, with an image style reference, a switchable single/dual sampler path, embedding-level seed variance, and an experimental SDE noise-injection stage.

---

# English

## 1. Required Custom Nodes

| Nodes used in this workflow | Package | Repository |

|---|---|---|

| TextEncodeKrea2OstrisEdit, Krea2OstrisEditModelPatch | ComfyUI-Krea2-Ostris-Edit | https://github.com/ostris/ComfyUI-Krea2-Ostris-Edit |

| ClownsharKSampler_Beta, ClownOptions_SDE_Beta | RES4LYF | https://github.com/ClownsharkBatwing/RES4LYF |

| LazySwitchKJ | ComfyUI-KJNodes | https://github.com/kijai/ComfyUI-KJNodes |

| Power Lora Loader, Image Comparer, Label | rgthree-comfy | https://github.com/rgthree/rgthree-comfy |

| RBG_Smart_Seed_Variance | ComfyUI-RBG-SmartSeedVariance | https://github.com/RamonGuthrie/ComfyUI-RBG-SmartSeedVariance |

| ColorCorrect | comfyui-art-venture | https://github.com/sipherxyz/comfyui-art-venture |

| AGSoft_Empty_Latent_Krea2 | comfyui-AGSoft | https://github.com/Art-xmaster/comfyui-AGSoft |

Core ComfyUI nodes (no install needed): UNETLoader, CLIPLoader, VAELoader, LoadImage, VAEDecodeTiled, SaveImage.

> ⚠️ **Heads-up on "Install Missing Nodes":** several nodes in this JSON carry an incorrect cnr_id comfyui_fearnworksnodes) in their saved metadata. ComfyUI Manager may therefore resolve them to the wrong package. If auto-install misbehaves, install the repos in the table above manually.

### Models expected

| Slot | File |

|---|---|

| UNET (diffusion model) | krea2_turbo_bf16.safetensors |

| CLIP / text encoder | qwen3VLInstruct4bHeretic_v10.safetensors — loader type set to *krea2** |

| VAE | wan_2.1_vae.safetensors |

| LoRAs (enabled) | Krea 2/style/krea2_style_reference.safetensors, Krea 2/style/Purple_Grainy_Kr2_AM.safetensors |

Any Krea 2–compatible Qwen3-VL text encoder will work; the CLIPLoader **type must stay krea2**. LoRA entries point at a Krea 2/style/ subfolder — either mirror that folder layout or re-pick the files in the loader.

## 2. How to Use

### Inputs (blue group, left)

- **Node 30** (green) — your **positive prompt**. The shipped prompt is a long descriptive paragraph; Krea 2 responds well to prose rather than tag lists.

- **Node 34** (red) — **negative prompt**, empty by default.

- **Node 39** AGSoft_Empty_Latent_Krea2 — output size. Set to mode Preset → Landscape - 1920x1080 (16:9). Switch the mode widget to use the manual width/height fields instead.

### Style reference

This workflow has **two independent style controls** — use either or both:

1. **Image reference** — LoadImage (node 31) feeds image1 on both text encoders. The Ostris Krea 2 Edit encoder folds the reference image into the conditioning, so composition and look carry over from the source image. Nodes 30/34 have image2 and image3 slots free if you want to stack more references.

2. **LoRA stack** — Power Lora Loader (node 16) holds 12 Krea 2 style LoRAs; two are on by default krea2_style_reference, Purple_Grainy_Kr2_AM), the rest are toggled off and ready to experiment with. Toggle with the switch on each row and adjust strength inline.

Docs: https://github.com/rgthree/rgthree-comfy

To run pure text-to-image with no image influence, disconnect or bypass node 31.

### The sampler switch (Lazy Switch KJ)

**Node 26** LazySwitchKJ picks which sampler path actually renders. The boolean widget is the only thing you touch:

| Switch | Path | Configuration |

|---|---|---|

| *false** (current) | **Single stage** → node 37 | 8 steps, exponential/ddim + beta57, denoise 1.0, CFG 1.0, seed randomized |

| *true** | **Two stage** → node 22 → node 36 | Stage 1 (node 22): 8 of 10 steps, exponential/ddim + beta57, denoise 1.0 · Stage 2 (node 36): final 2 steps, exponential/res_4s_munthe-kaas + kl_optimal, denoise 0.27 |

"Lazy" means the unused branch is never executed — flipping the switch costs nothing on the idle path. The two-stage route is a refiner pass: a fast DDIM block lays down composition, then a short high-order Munthe-Kaas pass at low denoise cleans up detail.

Sampler docs: https://github.com/ClownsharkBatwing/RES4LYF · Switch docs: https://github.com/kijai/ComfyUI-KJNodes

### Seed variance (node 40)

RBG_Smart_Seed_Variance sits between the positive encoder and all three samplers. It injects structured noise into the **text embeddings** rather than the latent, which produces genuine compositional variety from the same prompt without breaking prompt adherence — useful on Krea 2, where changing the sampler seed alone often barely moves the image.

Current settings: preset 🌱 Subtle, model type 📸 Krea2 (SingleStream), fade Instant, injection Beginning Steps, no token protection, no direction shift, schedule constant, seed randomized.

- Beginning Steps applies the variance over roughly the **first 20%** of the timeline — composition roams, detail rendering stays clean.

- Raise variance_preset (Subtle → Balanced → Creative → Bold → Wild) for more spread. Push too far and prompt adherence degrades.

- *model_type sets the strength multipliers** — Krea2 is 1.2× strength / 1.1× randomize against the Z-Image baseline. Across the full model table these span a 20× range, so the setting matters enormously on sensitive models; on Krea 2 the effect is mild.

- direction_shift changes the character of the noise — each option has its own generator. The node also contains Krea2-specific 12-band shaping logic, but it is gated behind an embedding width divisible by 12; the Qwen3-VL encoder used here is 2560-wide, so that path never activates. The direction shifts still work, just without the band multipliers layered on top.

- cutoff_step / total_steps are inert while variance_schedule is constant; they only matter for the hard_lock, step_cutoff, decreasing and tiered_release modes.

Docs: https://github.com/RamonGuthrie/ComfyUI-RBG-SmartSeedVariance

### Output and post-processing

1. **Node 17** VAEDecodeTiled — tiled decode (1024 tile / 128 overlap), kept because large decodes are memory-heavy on Macs. On a big-VRAM GPU a plain VAEDecode is faster.

2. **Node 23** ColorCorrect — the post-processing stage. Temperature / hue / brightness / contrast / saturation / gamma, all currently **neutral**, so it's a pass-through until you dial something in. This is the place for a final grade rather than baking it into the prompt.

Docs: https://github.com/sipherxyz/comfyui-art-venture

3. **Node 19** Image Comparer (rgthree) — slider comparison, **A = colour-corrected**, **B = raw decode**, so you can see exactly what the grade did.

4. **Node 21** SaveImage — prefix Krea2/1. Note it is **not** in normal Always mode in this file, so a fresh import will not auto-save; set it to Always if you want every run written to disk.

## 3. Node 25 — Experimental Noise Injection

**Node 25 ClownOptions_SDE_Beta is an experiment.** It feeds an OPTIONS bundle into all three ClownsharKSamplers (nodes 22, 36 and 37), adding SDE-style noise on top of the sampler's own eta — churning noise back in during sampling to push texture and grain rather than letting the image resolve smoothly.

Current settings: gaussian noise type (step and substep), hard noise mode, eta 0.5 and eta_substep 0.5, fixed seed.

**It is entirely optional.** Bypass it Ctrl+B) and the workflow runs fine on the samplers' built-in settings — you'll get a cleaner, smoother result. Tweak it, mute it, or delete it as you see fit. Lower the etas for a subtler effect; raise them for more grit. Treat it as a texture dial, not a required part of the pipeline.

## Quick Start

1. Install the seven packages in the table above, restart ComfyUI.

2. Load the three models; confirm CLIPLoader type is krea2.

3. Drop a reference image into node 31 (or bypass it for pure txt2img).

4. Write your prompt in node 30.

5. Set size in node 39.

6. Leave node 26 on false for a fast 8-step run; flip to true for the refined two-stage pass.

7. Optional: bypass node 25 if you want a cleaner image.

8. Queue.

---

---

# 中文 (简体)

Krea 2 Turbo 文生图工作流，包含图像风格参考、可切换的单/双采样器路径、嵌入层级的种子变化，以及一个实验性的 SDE 噪声注入环节。

## 1. 所需自定义节点

| 本工作流使用的节点 | 插件包 | 仓库地址 |

|---|---|---|

| TextEncodeKrea2OstrisEdit, Krea2OstrisEditModelPatch | ComfyUI-Krea2-Ostris-Edit | https://github.com/ostris/ComfyUI-Krea2-Ostris-Edit |

| ClownsharKSampler_Beta, ClownOptions_SDE_Beta | RES4LYF | https://github.com/ClownsharkBatwing/RES4LYF |

| LazySwitchKJ | ComfyUI-KJNodes | https://github.com/kijai/ComfyUI-KJNodes |

| Power Lora Loader, Image Comparer, Label | rgthree-comfy | https://github.com/rgthree/rgthree-comfy |

| RBG_Smart_Seed_Variance | ComfyUI-RBG-SmartSeedVariance | https://github.com/RamonGuthrie/ComfyUI-RBG-SmartSeedVariance |

| ColorCorrect | comfyui-art-venture | https://github.com/sipherxyz/comfyui-art-venture |

| AGSoft_Empty_Latent_Krea2 | comfyui-AGSoft | https://github.com/Art-xmaster/comfyui-AGSoft |

ComfyUI 核心节点（无需安装）UNETLoaderCLIPLoaderVAELoaderLoadImageVAEDecodeTiledSaveImage。

> ⚠️ **关于"安装缺失节点"的提醒：** 本 JSON 中有若干节点的元数据带有错误的 cnr_idcomfyui_fearnworksnodes），ComfyUI Manager 可能因此解析到错误的插件包。若自动安装出现异常，请手动安装上表中的仓库。

### 所需模型

| 类别 | 文件 |

|---|---|

| UNET（扩散模型） | krea2_turbo_bf16.safetensors |

| CLIP / 文本编码器 | qwen3VLInstruct4bHeretic_v10.safetensors —— 加载器类型设为 *krea2** |

| VAE | wan_2.1_vae.safetensors |

| LoRA（已启用） | Krea 2/style/krea2_style_reference.safetensorsKrea 2/style/Purple_Grainy_Kr2_AM.safetensors |

任何兼容 Krea 2 的 Qwen3-VL 文本编码器均可使用，但 CLIPLoader 的 **type 必须保持 krea2**。LoRA 路径指向 Krea 2/style/ 子目录，请保持相同的目录结构，或在加载器中重新选择文件。

## 2. 使用说明

### 输入区（左侧蓝色分组）

- **节点 30**（绿色）—— **正向提示词**。自带的提示词是一段较长的描述性文字；Krea 2 对散文式描述的响应优于标签堆砌。

- **节点 34**（红色）—— **负向提示词**，默认为空。

- **节点 39** AGSoft_Empty_Latent_Krea2 —— 输出尺寸。当前为 Preset 模式 → Landscape - 1920x1080 (16:9)。将模式切换后可改用手动宽高输入。

### 风格参考

本工作流有**两套独立的风格控制**，可任选其一或同时使用：

1. **图像参考** —— LoadImage（节点 31）同时接入两个文本编码器的 image1。Ostris Krea 2 Edit 编码器会将参考图融入 conditioning，因此构图与画面质感会从源图延续过来。节点 30/34 的 image2image3 插槽仍空置，可叠加更多参考图。

2. **LoRA 堆叠** —— Power Lora Loader（节点 16）内含 12 个 Krea 2 风格 LoRA，默认启用两个krea2_style_reference 与 Purple_Grainy_Kr2_AM），其余已关闭，可自行尝试。每行的开关可独立切换，强度可就地调整。

文档：https://github.com/rgthree/rgthree-comfy

若想进行不受图像影响的纯文生图，请断开或旁路节点 31。

### 采样器切换（Lazy Switch KJ）

**节点 26** LazySwitchKJ 决定实际执行哪条采样路径。你只需要操作那个布尔开关：

| 开关 | 路径 | 配置 |

|---|---|---|

| *false**（当前） | **单段** → 节点 37 | 8 步exponential/ddim + beta57，denoise 1.0，CFG 1.0，随机种子 |

| *true** | **两段** → 节点 22 → 节点 36 | 第一段（节点 22）：10 步中运行 8 步exponential/ddim + beta57，denoise 1.0 · 第二段（节点 36）：最后 2 步exponential/res_4s_munthe-kaas + kl_optimal，denoise 0.27 |

"Lazy（惰性）"意味着未被选中的分支根本不会执行，因此切换开关不会在闲置路径上产生任何开销。两段式路径本质是一次精修：先用快速的 DDIM 段确定构图，再用简短的高阶 Munthe-Kaas 低 denoise 段收拾细节。

采样器文档：https://github.com/ClownsharkBatwing/RES4LYF · 开关文档：https://github.com/kijai/ComfyUI-KJNodes

### 种子变化增强（节点 40）

RBG_Smart_Seed_Variance 位于正向编码器与三个采样器之间。它将结构化噪声注入**文本嵌入**而非潜空间，因此能在同一提示词下产生真正的构图多样性，同时不破坏提示词贴合度 —— 这对 Krea 2 尤其有用，因为仅更换采样种子往往几乎不改变画面。

当前设置：预设 🌱 Subtle、模型类型 📸 Krea2 (SingleStream)、衰减 Instant、注入 Beginning Steps、无 token 保护、无方向偏移、调度 constant、种子随机。

- Beginning Steps 会在时间线的**前约 20%** 施加变化 —— 构图自由发挥，细节渲染保持干净。

- 提高 variance_preset（Subtle → Balanced → Creative → Bold → Wild）可获得更大差异；过高会损害提示词贴合度。

- *model_type 决定强度倍率** —— 以 Z-Image 为基准Krea2 为 1.2 倍强度 / 1.1 倍随机化。在完整的模型列表中这些倍率相差可达 20 倍，因此该设置对敏感模型影响极大；但在 Krea 2 上差异较小。

- direction_shift 改变噪声的*性质* —— 每个选项都有各自的噪声生成器。该节点内还有一套针对 Krea2 的 12 频带整形逻辑，但其触发条件是嵌入维度能被 12 整除；本工作流使用的 Qwen3-VL 编码器维度为 2560，因此该逻辑永远不会激活。方向偏移本身仍然有效，只是不会叠加频带倍率。

- 当 variance_schedule 为 constant 时cutoff_step 与 total_steps 不起作用；它们仅在 hard_lockstep_cutoffdecreasingtiered_release 模式下生效。

文档：https://github.com/RamonGuthrie/ComfyUI-RBG-SmartSeedVariance

### 输出与后期处理

1. **节点 17** VAEDecodeTiled —— 分块解码（分块 1024 / 重叠 128）。保留该节点是因为在 Mac 上大尺寸解码非常吃内存；若使用大显存显卡，普通 VAEDecode 会更快。

2. **节点 23** ColorCorrect —— 后期处理环节。色温 / 色相 / 亮度 / 对比度 / 饱和度 / 伽马，目前全部为**中性值**，因此在你调整之前它只是直通。最终调色建议在此完成，而不是写进提示词。

文档：https://github.com/sipherxyz/comfyui-art-venture

3. **节点 19** Image Comparer (rgthree) —— 滑块对比，**A = 调色后**，**B = 原始解码**，可直观看到调色带来的差异。

4. **节点 21** SaveImage —— 前缀 Krea2/1。请注意它在本文件中**并非**处于常规的 Always 模式，因此刚导入时不会自动保存；如需每次运行都写入磁盘，请将其设为 Always。

## 3. 节点 25 —— 实验性噪声注入

**节点 25 ClownOptions_SDE_Beta 是一项实验。** 它向全部三个 ClownsharKSampler（节点 22、36、37）输入一个 OPTIONS 组合，在采样器自身的 eta 之上叠加 SDE 式噪声 —— 在采样过程中重新掺入噪声，以强化质感与颗粒，而不是让画面平滑收敛。

当前设置：噪声类型 gaussian（主步与子步）、噪声模式 hardeta 0.5eta_substep 0.5、固定种子。

**该节点完全可选。** 将其旁路Ctrl+B）后，工作流会依照采样器自身的设置正常运行，得到更干净、更平滑的结果。你可以随意调整、静音或删除它。降低 eta 可获得更细微的效果，提高则更粗粝。请把它当作一个质感调节旋钮，而非流程中的必需环节。

## 快速上手

1. 安装上表中的七个插件包，重启 ComfyUI。

2. 载入三个模型；确认 CLIPLoader 的类型为 krea2。

3. 在节点 31 放入参考图（或将其旁路以进行纯文生图）。

4. 在节点 30 写入提示词。

5. 在节点 39 设置尺寸。

6. 节点 26 保持 false 可快速运行 8 步；切换为 true 则执行两段式精修。

7. 可选：旁路节点 25 以获得更干净的画面。

8. 加入队列执行。

Files with SHA256: 71709b1c1bc7c258d3b8d6ddcb84c2ec4363641586763e9c0952578624daa0d5