ComfyUI-BawkNodes Introduction
ComfyUI-BawkNodes brings the main parts of a FLUX image workflow into a small set of ComfyUI nodes. You can load a model, choose and apply LoRAs, prepare a prompt, generate or edit an image, and save the result without assembling each stage from many separate nodes.
The extension supports FLUX.1 dev, Krea, Kontext, FLUX.2 dev, and FLUX.2 klein 4B and 9B. It can also show compatible GGUF model files in its dropdowns when ComfyUI-GGUF is available.
It is especially useful when you want to compare LoRA combinations, add variety to prompts, work from an existing image, or generate images from a list of prompts. The author also provides example workflows for the supported model families, so you can see how the nodes connect.
How ComfyUI-BawkNodes Works
Think of a workflow as a path your image takes from an idea to a saved file:
Bawk Model Loader ──────────────────┬─► Bawk Wildcard Encoder ─► Bawk Sampler ─► Bawk Image Saver
Bawk LoRA Stack (Set A) ─► Bawk LoRA Switch ─┘
Bawk LoRA Stack (Set B) ─┘
- Choose the ingredients. Bawk Model Loader brings together the FLUX diffusion model, its text encoder or encoders, and its VAE.
- Choose a style or subject adjustment. A Bawk LoRA Stack holds up to six LoRAs. You can connect several stacks or use Bawk LoRA Switch to select one set.
- Prepare the prompt. Bawk Wildcard Encoder applies the selected LoRAs to the model and text encoder. It can add their trigger words and resolve choices such as
{red|green|blue}into a finished prompt. - Make the image. Bawk Sampler handles sampling and decoding. Depending on the model and inputs you choose, it can generate from text, work from an input image, or use reference images for editing.
- Keep the result. Bawk Image Saver saves the image and can retain information that helps you revisit the workflow.
There is no negative-prompt input in this node pack because its FLUX workflows do not use negative prompts. One exception worth distinguishing is the sampler’s CFG setting for klein base models: those models use real CFG against an empty negative, rather than asking you to write a negative prompt.
ComfyUI-BawkNodes Features
Bawk Model Loader
This node loads the diffusion model, VAE, and text encoder or encoders together. These are the main components the workflow needs to interpret your prompt and produce an image. Select files that match the model family you are using; the model table below shows the combinations listed by the author.
If ComfyUI-GGUF is installed, .gguf diffusion models and text encoders appear in the same dropdowns as other compatible files.
Bawk LoRA Stack
A LoRA adjusts a model’s output, often to introduce a particular style, character, or visual concept. Each Bawk LoRA Stack has six slots. For each slot, you can select a LoRA, set its strength, and enter trigger words.
- Strength controls whether a LoRA is active and how strongly it is applied. At a strength of
0, its trigger words are not added to the prompt. - Trigger words travel with their LoRA. When you change the active set, the words used by that set change too.
enabledturns an entire stack off.lora_stacklets you chain stacks when you want to combine more than one set.
The node also provides lora_triggers and lora_info strings. For example, you could keep one stack for a painterly look and another for a different visual direction, then switch between them without manually editing the associated words each time.
Bawk LoRA Switch
Bawk LoRA Switch accepts up to four sets, labeled set_a through set_d. Its active dropdown selects which set is loaded. Connect its output to the Wildcard Encoder’s lora_stack input.
This makes comparisons easier: you can prepare several LoRA combinations and change the selected set from one control instead of rebuilding the workflow.
Bawk Wildcard Encoder
The encoder applies the incoming LoRA stack to both MODEL and CLIP, then prepares the text used for generation.
trigger_wordscan place active LoRA trigger words at the start of your prompt, at the end, or leave them out. Words already present in the prompt are not repeated.- Wildcards let one prompt contain alternatives. For example,
a {red|green|blue} coatresolves to one of those colors. Choices can be nested. wildcard_seedcontrols the wildcard selection, helping you repeat a particular resolved prompt when needed.
The node outputs the resolved prompt, so you can see the actual wording used, and a LORA_INFO summary such as name@strength for the saver.
Bawk Sampler
Bawk Sampler combines latent preparation, sampling, and VAE decoding. It detects whether the selected model belongs to FLUX.1 or FLUX.2 and chooses the corresponding latent layout and schedule.
| Setting | What it does |
|---|---|
resolution | Choose a preset, Match input / reference image, or Custom. A linked WIDTHxHEIGHT string also works. The available presets include 4:5 and 5:4. |
steps | Sets the number of sampling steps. The author lists 20 for FLUX.1 dev, Kontext, and FLUX.2 dev, and 4 for klein distilled. |
guidance | Controls prompt guidance where the model uses it. The author lists 3.5 for FLUX.1 dev, 2.5 for Kontext, and 4.0 for FLUX.2 dev. It is ignored for klein distilled. |
scheduler | With auto, uses simple for FLUX.1 and the official FLUX.2 schedule for FLUX.2. |
input_image and denoise | Enable image-to-image work: start from an existing image and control the denoising used during generation. |
reference_image and reference_image_2 | Supply reference images for Kontext and FLUX.2/klein editing. The node scales them to the model’s preferred size and attaches them as reference latents. |
cfg | Leave at 1.0 for distilled models. Klein base models use real CFG, approximately 4–5, against an empty negative. |
max_shift and base_shift | FLUX.1-only controls. |
These values are guidance from the original README, not a requirement that every image use the same settings. For example, choose Match input / reference image when you want the output size to follow an image you supplied, or Custom when you have a specific size in mind.
Bawk Image Saver
The saver writes images under output/[model]-DD-MM-YYYY/, grouping results by model and date.
- PNG files embed the workflow. You can drag one back into ComfyUI to reopen it.
- JPG and WebP files receive a separate
.jsonsidecar for workflow information. - An optional prompt
.txtfile can include the LoRA list. - An optional Discord webhook can upload the result.
Pass the encoder’s LORA_INFO output to the saver when you want the saved information to include the LoRAs used.
Bawk Image Loader and Bawk Batch Processor
Bawk Image Loader loads an image, offers resize and pad-to-square options, and outputs its dimensions. This is useful when an existing image needs preparation before editing or sampling.
Bawk Batch Processor reads one row at a time from a CSV or JSON file. Supported row information includes prompt, seed, steps, guidance, and resolution. Set batch_index to increment to walk through a list of prompts across runs. This can help when you have several ideas to render with a consistent workflow.
ComfyUI-BawkNodes Models
Bawk Model Loader supports the following model families. The file names shown are the encoder and VAE selections identified in the original README; choose files appropriate to the model you are loading.
| Model family | clip_name1 | clip_name2 | vae_name | When you might use it |
|---|---|---|---|---|
| FLUX.1 dev / Krea / Kontext | t5xxl_*.safetensors | clip_l.safetensors | ae.safetensors | Use the matching workflow for the FLUX.1-family model you selected; Kontext supports the reference-image editing inputs described above. |
| FLUX.2 dev | mistral_3_small_flux2_*.safetensors | none | flux2-vae.safetensors | Use for a FLUX.2 dev workflow with its single text encoder. |
| FLUX.2 klein 4B | qwen_3_4b.safetensors | none | flux2-vae.safetensors | Use with the klein 4B workflow. The author provides a text-to-image example. |
| FLUX.2 klein 9B | qwen_3_8b_*.safetensors | none | flux2-vae.safetensors | Use with the klein 9B workflow. The author provides a two-reference editing example. |
The author’s workflow examples cover FLUX.1 dev text-to-image, LoRA sets with trigger words and a switch in flux1_dev_lora_sets.json, image-to-image, Kontext editing, FLUX.2 dev, klein 4B text-to-image, and klein 9B two-reference editing. These examples are a useful starting point when you are unsure which inputs belong together.
What's New with ComfyUI-BawkNodes
Version 3.1.0
- Trigger words per LoRA slot: The prompt can follow whichever LoRAs are active without you having to maintain a separate list of words.
- Bawk LoRA Switch: Prepare sets A through D and select the one you want to use.
LORA_INFOoutput and saver input: Keep a clearer record of the LoRAs used for an image.- 4:5 and 5:4 resolution presets: Choose these aspect ratios directly from the sampler.
Version 3.0.0
- FLUX.2 dev and klein support: Added the latent layout, schedule, and single-text-encoder handling needed for these model families.
- GGUF loading and reference images: Added GGUF file selection through ComfyUI-GGUF and reference-image handling for Kontext and klein editing.
- Guidance, metadata, and validation fixes: Corrected guidance that had not been applied, PNG metadata that could not be loaded back into ComfyUI, and disabled image-loader validation.
- Reworked LoRA sets: A chainable, toggleable Bawk LoRA Stack replaced the earlier dynamic stack and apply pair.
- Removed unused or placeholder components: The Z-Image loader, placeholder ControlNet node, unused AI prompt enhancement code, and unused utilities were removed.
- Changed widget layouts: Workflows saved with version 2.x need their Bawk nodes re-added before use with version 3.x.
Troubleshooting ComfyUI-BawkNodes
A saved version 2.x workflow no longer works as expected
- Open the workflow in ComfyUI.
- Re-add its Bawk nodes using the current node versions.
- Reconnect them, using a current example workflow as a guide.
- Check the selected model files and settings before generating again.
Version 3.0.0 changed widget layouts, so older saved Bawk nodes need to be replaced.
A .gguf file does not appear in a model dropdown
- Check whether ComfyUI-GGUF is available in your ComfyUI setup.
- Return to Bawk Model Loader and check the dropdown again.
GGUF files appear in the loader’s dropdowns when ComfyUI-GGUF is installed.
A LoRA’s trigger words are missing from the prompt
- Check that the LoRA’s strength is not
0. - Check that its stack is
enabled. - If you are using Bawk LoRA Switch, confirm that its
activeselection is the set containing that LoRA. - In Bawk Wildcard Encoder, set
trigger_wordstoprependorappend, rather thanoff. - Inspect the encoder’s resolved prompt. If the words were already in your original prompt, the encoder will not repeat them.
The same wildcard prompt produces different wording
Check wildcard_seed in Bawk Wildcard Encoder. Wildcards select from the alternatives you write inside braces; the resolved prompt output shows which choice was used for that run.
Dragging a saved image into ComfyUI does not reopen the workflow
Check the file type. The saver embeds the workflow in PNG files, which you can drag back into ComfyUI. For JPG and WebP, look for the accompanying .json sidecar instead.
Frequently Asked Questions
Do I need to write a negative prompt?
No. This pack does not provide a negative-prompt input for its FLUX workflows.
Why is guidance not changing my klein distilled result?
The author notes that guidance is ignored for klein distilled. Its listed step count is 4, and cfg should remain at 1.0 for distilled models.
Can I keep more than six LoRAs in a workflow?
Yes. Each Bawk LoRA Stack has six slots, and stacks can be chained through lora_stack. You can also prepare different sets for the LoRA Switch.
Can I edit an existing image?
Yes. Use input_image with denoise for image-to-image work. Kontext and FLUX.2/klein editing workflows can also use the sampler’s reference-image inputs, as appropriate for the selected model.
Learn More about ComfyUI-BawkNodes
- Example workflows: See complete examples for each supported model family and for image-to-image and reference-image editing.
- LoRA sets and switch example: See how trigger words, LoRA stacks, and the switch connect in a FLUX.1 dev workflow.
- ComfyUI: Learn about the application in which these nodes run.
- ComfyUI-GGUF: Read about the additional component used to make GGUF files available in the model dropdowns.
