Laziness is the Engine of Inference: How I Defeated the Showcase Nightmare and Created the PromptForge Ecosystem for ComfyUI
What is PromptForge?
PromptForge is an autonomous visual knowledge base and environment orchestrator for your generations, working in tandem with ComfyUI. Forget about endless text notebooks, the chaos of hundreds of custom nodes, and monstrous Excel spreadsheets where you used to bury descriptions of characters, outfits, and working prompts for years.
This tool turns routine inference into an exciting game or a powerful automated conveyor. It has its own memory, history, visual previews, and can manage your LoRA models on the fly. You save hours of time by assembling complex scenes in just a few clicks.
Background: Why I Created It
Training LoRA models is pure joy. But formatting the final release on Civitai... When you train a single character, there’s no problem: you copy the template once and keep spinning images. But when I finished my massive multi-character LoRA for the KonoSuba universe, a real hell began.
Ten unique characters, canonical clothes, alternative outfits, a bunch of locations. To make a basic representative showcase for the model page, I experienced tons of stress. The process looked like this: minimize the window, copy the anatomy from here, paste it there, remember the clothing tags, come up with a scenario, set up the background. Prompting alone took me about 40 minutes! At that point, the celebratory release of the model turns into a dull torment.
I got tired of it. Having no specialized education in programming (my maximum in school was limited to three-dimensional matrices and basic sorting), I turned on the "vibe-coding" mode and, together with Gemini and Claude, built PromptForge for my tasks.
At first, everything was simple: I wanted to make a constructor where you could forget about manually entering the anatomical basis of characters and assemble prompts with clicks. Done. But laziness progressed. Why should I copy text by hand every time? Let the program send data directly to ComfyUI. Done. Further more: why should I manually rearrange LoRA loader nodes on the canvas when changing a character? Let the necessary LoRA connect itself when activating a character card or their custom outfit, and let the extra ones disconnect.
After 30 hours of building strict logic and catching bugs, I realized that I no longer needed to keep the Comfy tab open at all and break my fingers over Alt+Tab on a single monitor. A full-fledged inference management ecosystem is ready.
Connection example:
Anatomy of PromptForge: What the Software Consists Of
The application interface is divided into four main panels, each covering its own piece of the overall pipeline.
1. Builder (Scene Constructor and Generation)
This is your main working screen and command center. This is where the assembly magic happens: you select the necessary inputs (characters, clothes, scenarios, styles), and they are dynamically activated inside the prompt template. By default, a template is built-in that covers 99% of everyday tasks, but you can easily create your own templates for unique needs - no rigid restrictions.
Generation parameters, negative prompt, connection to ComfyUI, and model management are also configured here.
Important: For the auto-binding magic to work, you will need my custom nodes for ComfyUI (PromptForge-Nodes) - a connector and a custom lora loader. I have already attached a ready-made working workflow on GitHub. The connection is laughably simple - just two nodes are added to the canvas.
2. Library (Your Knowledge Base and Concept Storage)
Library is your stationary data storage and source of inspiration. Here you can walk through subfolders, scroll through visual references, add new entities, or delete unneeded ones. You fill it out once and use it for the rest of your life.
To help you understand how it works right away, you can download my ready-made starter library tailored for the Anima model. I did a massive amount of work and packed a huge array of data there: 220 characters, 84 outfits, 76 scenarios, and 44 unique styles.
I’m rolling out two important disclaimers right away:
There is a lot of NSFW and Ecchi there: Corresponding scenarios, poses, and clothing variants are included in the pack by default.
Model authorship and compatibility: A huge number of characters from third-party creators are configured in the database. Physically, I can only vouch for the ecosystem's operation and correct rendering on those models where the author is indicated as MonsieurBaka (my nickname on Civitai). All links to download third-party models are embedded directly into the cards - download them directly from the authors, I am not going to steal anyone else's clicks.
How to launch the ecosystem in a couple of steps:
Download the necessary LoRAs from links in Library and drop them into your standard models/loras folder inside ComfyUI.
For reliability, restart Comfy and PromptForge.
Connect ComfyUI in the Builder tab, go to Library and click the Check LoRA dependencies button.
The program will scan your folders itself and tightly bind the downloaded .safetensors files to the text names of characters, styles, and clothes.
3. Tools (Auxiliary Tools)
This category in the library stands apart. Remember how often when working, for example, with the Klein 9B architecture, you got three hands and three legs in the frame? It happened regularly in 100% of cases. To solve such problems, things like Anatomy Fixer are created. This is not a visual style, it doesn't have its own unique triggers - it's a technical utility. Such tools are moved to the Tools panel. They are not required, but they are irreplaceable when you need to precisely back up anatomy.
4. History & Gallery (History and Verification)
A local archive of your generations. Here you can track the history of sent prompts and view the resulting images right inside the application, quickly evaluating the results without jumping to third-party viewers.
On "Wild" Inference and Model Compatibility
Let me clarify an important point: hundreds of characters from a wide variety of authors are configured in my starter library. But I physically cannot guarantee that absolutely every third-party LoRA from the internet will give you an ideal result on the first try. And when you deploy the database on your machine, you will clearly see which models are built to last and which ones begin to fall apart.
When I was assembling and filling this ecosystem, I faced a monumental task - to generate more than 700 unique images just for preview cards! Naturally, I had neither the time nor the desire to engage in manual selection of lucky seeds (cherry-picking), look for favorable angles, or spend hours adjusting weights for each lonely model. Absolutely everything spun on a random seed in a tight streaming conveyor mode.
To verify all 220 characters and 84 outfits, I used the same basic, maximally "sterile" scenario:
solo, frontal, neutral pose standing, hands on sides, white background, even lightingThis is the perfect litmus test for any model. If a LoRA is trained according to all the rules of modern dataset hygiene, it will easily give you a clean character on a flat white background with studio lighting.
But if, when activating this scenario, instead of a white background you get pitch darkness, terrible dramatic shadows, or twisted anatomy, this is the main indicator that the third-party model is overtrained. The author simply over-cranked the number of steps (for some reason setting 3000–6000 steps where modern architectures only need 800–1500), causing the character concept to merge tightly with the lighting and background from the original frames.
In such situations, you will inevitably have to "play with weights" - manually lower the strength of the third-party LoRA (for example, to 0.6–0.7) or try to overpower its built-in background with more aggressive prompting. My personal models pass this test in the builder without a hitch, but I bear no responsibility for the flexibility of the rest of the Civitai zoo.
A small reminder for authors who train LoRAs:
If you want your models to remain plastic and easily integrate into such visual knowledge bases, please do not forget to specify basic environment variables in captions:
Type and scale of the shot (frontal shot, medium shot, full body portrait)
Background at least in a couple of words (plain background, cityscape, indoor).
Lighting, even in its simplest form (day, night, even lighting).
Two words in the dataset layout free the character's pixels from the captivity of the original context and allow the model to adequately respond to any custom scenarios.
Pipeline in Action: From Click to Finished Image
When the base is filled and linked to models, generating a complex scene turns into a casual conveyor:
You open the Builder tab and click to activate the necessary inputs saved in your database: Character (for example, Aqua) -> Clothes -> Scenario -> STYLE.
Click Generate Prompt. The program instantly assembles the text array. If LoRA models are bound to the selected cards, they will automatically load into the stack in ComfyUI.
Stage of hybrid control: Before sending to Comfy, look at the resulting text in the manual editing window. For example, you chose the "close-up selfie" scenario, but Aqua's canonical clothing template specifies her boots. Why overload the model's context window with tokens if the shoes physically won't fit in the frame? Right here, cut out the extra boots by hand, make the smile pretty, and tweak small details.
Click Generate in ComfyUI and pick up the finished image from the gallery. No chaos on the canvas, no copy-pasting.
Manifesto of Openness
From now on, I am introducing a new rule for myself: when developing and publishing each of my major LoRAs (now, by the way, a huge mega-dataset for Mushoku Tensei is in the compilation process), I will attach a separate link to a ready-compiled data folder for PromptForge. You just download the archive, drop the files into your local program data directory - and you instantly have a ready-made universe deployed with all previews, clothing blanks, and configured weight connections. I encourage other authors to pick up this initiative as well. Let's share not just model "blanks", but ready-made working inference ecosystems.
The program is completely free, the source code is open and available for any modifications. You have every right to rewrite it for your needs, integrate it into any studio pipelines or, if you want, completely redesign the interface, covering it with skins of Miku or Teto and adding voice acting to the buttons, xD.
Modern AI assistants have blurred the lines. If you are uncomfortable working in existing interfaces, create your own, automate the routine, and share the results with the community.
Download PromptForge:
Library for Anima on HuggingFace .zip
A Step Back: How I Saved My Datasets with TextManager
Before you even get to the inference stage, you have to go through purgatory - dataset preparation and labeling. And if for one simple model the captions can be sorted out by hand, on large-scale projects assembling the right tags turns into a separate kind of hard labor. At this stage, I tried existing software and realized: it is either hopelessly outdated or tries to be smarter than the author.
Usually, utilities like img-txt_viewer are used to view images and text descriptions. But when your dataset grows to hundreds of images, the working logic of such analogues begins to frankly harm the pipeline:
Self-willed punctuation: Such programs love to automatically and permanently replace all periods with commas. For modern DiT architectures, which are critically tied to pure Natural Language and strict semantics, this is a critical error that completely breaks the description syntax.
“On-the-fly editing” without insurance: Changes to files are overwritten on the fly right during input. One wrong move, an accidental click - and you have irreversibly ruined the text file layout without the possibility of undoing it.
Complete lack of batch processing: If midway through you realize that some tag, aspect of the description, or synthetic trigger needs to be mass deleted, replaced, or added across all 500 files - you are doomed to monotonous manual work.
I got tired of tolerating this, so I wrote TextManager - an autonomous open-source utility for bringing total order to text descriptions (.txt) for your images.
What TextManager can do:
Killer feature - a full-fledged Batch Mode (Batch mode): The main tool for routine automation. Allows you to filter, mass replace, delete, or add the necessary tokens and synthetic triggers across the entire database in a couple of clicks.
Isolated saving as fool-proofing: The program works carefully and writes changes to .txt files only by your explicit command, providing ironclad protection against accidental errors and crooked overwrites during operation.
Absolute syntax control: No self-willed auto-replacements of punctuation marks. The program strictly maintains the punctuation you specify, allowing you to build complex text structures for training.
Download TextManager:
TextManager on Github
P.S. PromptForge Neo Armstrong Cyclone Jet Armstrong Cannon Pro Ultra Mega Version is in work. Better UI, better architecture, new features. Wait for the announcement.
Or create your own PromptForge Neo Armstrong Cyclone Jet Armstrong Cannon Pro Ultra Mega Version by YOU.
Actually in work 3.07.2026
I really love the new UI for the app. Screens for guys who've read till the end
