A compact ComfyUI workflow for Krea 2 text-to-image generation, with optional image refinement, LoRA support, and a comparison view. The main path keeps generation and refinement on the same diffusion model to balance quality, speed, and model loading overhead.
Workflow: Krea2_V2.2.json
The supplied V2.2 export also contains a separate SeedVR2 Upscale branch, muted by default. This branch uses additional models; it is distinct from the lightweight Krea 2 refinement path described below.
- Krea 2 text-to-image generation.
- Optional 1.5× image resize followed by low-denoise refinement with Krea 2.
- An Image Comparer connected to the base image and the Krea-refined image.
- Two Power LoRA Loader nodes for individual LoRAs or stacks, configured separately for generation and refinement.
- Per-layer conditioning control through ConditioningKrea2Rebalance on the generation path.
- An Options panel using Fast Groups Muter to enable or mute workflow groups.
- PNG output through SaveImageWithMetaData.
The active Upscale group uses this sequence:
Krea 2 generation → VAE decode → ImageScaleBy (bilinear, 1.5×)
→ VAE encode → Krea 2 refinement (denoise 0.15) → VAE decode
At the saved base resolution of 1280 × 1728, the resize produces 1920 × 2592 before refinement. Both samplers receive the diffusion model from the same Initial Loader, through their own LoRA loaders. No additional diffusion checkpoint is needed for this path.
This is a quality/speed tradeoff. It is not a claim of maximum reconstruction quality or superiority to dedicated upscalers such as SeedVR / SeedVR2.
The separate SeedVR2 Upscale branch receives the base decoded image, rather than the output of the Krea refinement pass. Its model loaders, upscaler, and save node are muted in the export (mode: 2). A connected FastFilmGrain node is bypassed (mode: 4). The comparer is not connected to SeedVR2 output.
Use a ComfyUI installation that supports Krea 2, the krea2 text-encoder type, subgraphs, and the er_sde sampler. The export contains an Initial Loader subgraph, not a missing custom node with a UUID-like name.
The JSON records frontend version 1.52.7 and mixed per-node core version metadata. These are export details, not a verified minimum-version requirement or a reproducible environment lock.
The main Krea 2 path uses .safetensors files with the standard ComfyUI loaders. It does not contain GGUF loaders.
| Component | Model or selection guidance | Default location |
|---|---|---|
| Krea 2 diffusion model | Any compatible Krea 2 .safetensors model of your choice |
ComfyUI/models/diffusion_models/ |
| Qwen3-VL 4B text encoder | qwen3vl_4b_fp8_scaled.safetensors |
ComfyUI/models/text_encoders/ |
| Qwen Image VAE | qwen_image_vae.safetensors |
ComfyUI/models/vae/ |
| Optional Krea 2 LoRAs | A local Iris file is enabled in both saved loaders; see below | ComfyUI/models/loras/ |
The author also uses krea2TurboFP8 (approximately 12 GB), a Qwen3-VL 4B encoder variant (approximately 5 GB), and the Qwen Image VAE (approximately 250 MB). A ComfyOrg nvfp4 Krea 2 variant (approximately 7 GB) is another reported option. These are approximate file sizes supplied by the author, not VRAM requirements; the nvfp4 variant is not selected in this export or validated here.
Select your preferred Krea 2 model in Initial Loader. Any Krea 2 model compatible with the workflow's standard .safetensors loader can be used. Saved model selections reflect local testing; choose the file installed on your machine without renaming it to match the export.
Model weights are not included in this repository. Obtain compatible weights from their publishers and follow their model cards. Different quantizations and model variants may need different runtime support and settings.
Anyone downloading or using Krea 2 weights must read and comply with the Krea 2 Community License Agreement (PDF). The model weights are governed by that agreement separately from this repository's MIT license.
| Package | Nodes used | Role in this export |
|---|---|---|
| rgthree-comfy | Power Lora Loader (rgthree), Fast Groups Muter (rgthree), Image Comparer (rgthree) |
LoRA loading, group controls, comparison |
| ComfyUI-Easy-Use | easy setNode, easy getNode |
Shared MODEL, CLIP, and VAE routing |
| nova452's conditioning nodes, now documented as Rebalance-Pack | ConditioningKrea2Rebalance |
Generation conditioning adjustment |
| Santodan's revived metadata extension | SaveImageWithMetaData |
Image saving with metadata |
| SeedVR / SeedVR2 (official project), via the ComfyUI integration | SeedVR2LoadDiTModel, SeedVR2LoadVAEModel, SeedVR2VideoUpscaler |
Additional branch, muted by default |
| VRGDG nodes recorded in the export, with utility nodes now in VRGDG Custom Node Pack | FastFilmGrain |
Bypassed node after SeedVR2 |
SeedVR2 and FastFilmGrain are additional dependencies present in V2.2 beyond the original four-package list. Their processing is not needed for the main path, but their node definitions may still be needed to load the complete canvas without missing-node placeholders.
Some saved registry identifiers are inconsistent with node names: the two Power LoRA loaders and the conditioning node carry cnr_id: comfyui-workflow-encrypt; FastFilmGrain carries cnr_id: comfyui-aspect-ratio-crop-node. Use the actual node types and upstream sources when resolving dependencies. These identifiers alone do not establish a need for a workflow-encryption package.
The metadata extension's upstream README currently points to a maintained fork. The table identifies the provider recorded in this workflow; compatibility with replacement forks has not been tested. Install only one provider for a given node type.
- Install or update ComfyUI following its upstream instructions, including the appropriate PyTorch/GPU setup for your hardware.
- Install the custom node packages above using ComfyUI-Manager, where available, or each package's documented manual installation procedure. For manual installations, use ComfyUI's own Python environment for any package requirements. Restart ComfyUI and refresh its interface afterward.
- Place your diffusion model, text encoder, VAE, and optional LoRAs in the model directories listed above, or use your existing ComfyUI model-path configuration.
- Download Krea2_V2.2.json and open it in ComfyUI by dragging it onto the canvas or using the workflow open command.
- Resolve any missing node types before queueing. Check the dependency table if automatic detection suggests an unrelated package.
- In
Initial Loader, select the files actually installed on your machine. Inside the subgraph, theCLIPLoadertype is saved askrea2, with devicedefault; theUNETLoaderweight dtype isdefault. - Review both Power LoRA loaders. Each export row is enabled at strength 1.0 and refers to
My_LoRa/K2_Iris_KSC_9.safetensors. Replace that selection with a compatible local LoRA, or switch the row off in both loaders for a run without LoRAs. The local file is not bundled or required.
Make these runtime selections in your own ComfyUI session or local workflow copy. Keep the repository's original JSON as the reference export.
- Write your positive prompt in the
Generationgroup's text encoder (node 500). Its saved text is simplywoman; the base negative prompt (node 505) is empty. - Set width, height, and batch size in
EmptyLatentImage(node 508). Start with the saved settings below and reduce resolution if memory is insufficient. - Choose the processing path using the Options group controls:
- Base generation only: keep
Krea2 MODEL,Base model,Generation, andSave Imageactive; muteUpscaleand keepSeedVR2 Upscalemuted. Also mute the separate Image Comparer node (32), which expects the refinement output on its second input. - Krea refinement: enable
Upscaleand the Image Comparer. KeepSeedVR2 Upscalemuted. This is the active upscale path in the saved export.
- Base generation only: keep
- Review the refinement prompt (node 18) and its Power LoRA Loader (node 536). This prompt is empty in the export and is independent of the generation prompt. The refinement negative conditioning is produced by
ConditioningZeroOut; the generation rebalancing node does not feed this pass. - Queue a generation. For refinement runs, inspect the base preview and the Image Comparer: A receives the base image and B receives the saved Krea-refined image.
- Find the PNG outputs under ComfyUI's configured output directory:
- Base images:
Krea2/ - Krea-refined images:
Krea2/Krea2_Upscale/
- Base images:
The save nodes use timestamp-based prefixes and metadata_scope: full. The comparer stores temporary image references in the JSON; these are not example images distributed with the repository. Run the workflow to populate it on your machine.
Keep this branch muted for the lightweight workflow. The official model project is SeedVR; the workflow's nodes are provided by the separate ComfyUI integration. If you choose to use this branch, follow that integration's installation and hardware guidance first. The saved branch selects:
- DiT:
seedvr2_ema_7b_sharp_fp16.safetensors - VAE:
ema_vae_fp16.safetensors - Device:
cuda:0, with CPU offloading - Resolution:
2000; maximum resolution:2592
Its upstream documentation uses ComfyUI/models/SEEDVR2/ for these models. This is an additional model set, separate from the Krea model and Qwen VAE. The branch saves under Krea2/Krea2_SeedVR/; its timing, memory use, and image quality are not established by the performance figures below. Preserve the FastFilmGrain bypass unless you intentionally want that processing.
These values describe the supplied JSON, not universal recommendations for every Krea 2 model or LoRA.
| Parameter | Base generation | Krea refinement |
|---|---|---|
| Sampler node | 504 |
19 |
| Steps | 8 |
2 |
| CFG | 1 |
1 |
| Sampler | euler |
er_sde |
| Scheduler | beta |
simple |
| Denoise | 1.0 |
0.15 |
| Saved seed | 6439 |
5047 |
| Seed control | increment |
increment |
| Positive prompt | woman |
Empty |
| Negative conditioning | Empty text encoding | Zeroed conditioning |
Other relevant settings:
- Base latent: 1280 × 1728, batch size 1.
- Resize: bilinear, factor 1.5.
- Generation conditioning multiplier: 0.5.
- Generation per-layer weights:
1.0,1.0,1.0,1.0,1.0,1.0,1.0,2.5,5.0,1.1,4.0,1.0. - Both saved LoRA rows: enabled, strength 1.0, configured independently.
For repeatable comparisons, fix the seeds in both samplers and keep the models, LoRAs, prompts, resolution, and other settings unchanged. The saved increment controls change seeds between runs. If you change LoRAs for generation, review the refinement loader too: its MODEL/CLIP inputs come from the shared initial models, not from the generation LoRA loader.
Author-observed timings on an NVIDIA RTX 5070 Ti with 16 GB VRAM and 48 GB system RAM:
| Path | Approximate observed time |
|---|---|
| Base generation | 29 seconds |
| Generation with the Krea upscale/refinement path | 60 seconds total |
These are practical reference observations, not universal benchmarks or measurements reproduced for this documentation. The observations do not establish timings for the exact model selection in the current export, and do not cover SeedVR2. Model variant, resolution, LoRA stack, software versions, offloading, and model-loading state can affect runtime and memory use.
Iris V2 — Krea 2 Character LoRA is an optional Krea 2 character LoRA published by the workflow author. It is not required for generation.
Consult its model card for the supported base model, trigger words, and recommended strength before selecting it in the loaders. The export's local K2_Iris_KSC_9.safetensors filename is not proof that it matches a particular public Iris V2 file. The model card and weights were not independently inspected during this documentation pass; end-to-end compatibility has not been tested here.
These screenshots show the canvas layout and controls. Model and LoRA selections shown in the interface may differ from the saved JSON; select your own compatible files as described above.
Overview of the workflow groups and their connections, including the separate SeedVR2 branch.
The generation section brings together model selection, LoRA loading, conditioning, sampler settings, and the Options group controls.
The Krea path uses a 1.5× bilinear resize and low-denoise refinement. The separate SeedVR2 branch uses its own DiT and VAE models and is muted by default in the supplied JSON.
Generated image examples have not been added yet. When adding examples, record the model and quantization, LoRA names and strengths, prompts, both seeds, resolution, and relevant sampler settings. Label base, Krea-refined, and any future SeedVR2 examples separately.
The workflow and documentation in this repository are released under the MIT License, copyright © 2026 NymZero. Model weights and third-party custom nodes remain subject to their respective licenses.
This documentation was checked against the nodes, connections, subgraph, and saved settings in Krea2_V2.2.json, with upstream documentation/source checks for the listed node providers. The workflow JSON was left unchanged. No ComfyUI generation or GPU performance test was run as part of this documentation update.


