1. Introduction

Hello everyone!

This is my first Manual for getting the best out of Chroma, After some times after full Chroma release, I decided to create my first guide about its general usage and share some tips for getting the best results.

⚠️ Disclaimer: I’m still a beginner myself, but I believe I understand the model well enough to generate decent outputs. I strongly encourage everyone to point out mistakes, missing details, or improvements.⚠️

📌 Sources used:

• Chroma CivitAI page

• Chroma HuggingFace page

• Official Discord

• Reddit threads and discussions

2. What is Chroma?

Chroma is a standalone 8.9B parameter model based on Flux.1-Schnell.

Key advantages:

• Fully uncensored (no NSFW/gore restrictions)

• Apache 2.0 license → free usage, images are 100% yours

• Strong prompt adherence compared to many other models ( no more SDXL randomness ! )

• Works with some Flux tools (ControlNet, LoRAs, etc.)

• Great results in both realism and stylized prompts (even without LoRA)

• Multiple versions available:

  • Chroma 1-HD Flash → optimized for speed

  • Chroma 1-Radiance → WIP pixel-space model, avoids VAE compression artifacts

👉 Think of Chroma as Flux without censorship, with much more freedom in both realism and artistic prompts.

3. Requirements

Personally i generate images locally, using my good ol' 3060 (god bless her🙏)
Recommended setup (ComfyUI example):

• Chroma 1-HD ( There are GGUF or FP8 versions)

• Flux VAE

• Text Encoder: After some experiments i don't recommend using default T5-XXL encoder. Personally i am using Flan-T5-XXL, but there is also GNER version, use whatever you like

• Workflow: I am using workflow made by Tenofas
  (quick tip: if you using some new Loras i encourage you to use workflows provide by the LoRA creator, In this way you can reach best result. For exaple i use Workflows made by TijuanaSlumlord to get best results with his loras)

3.1 Base settings

⚠️ Keep in mind these are the settings that work in my workflow.: schedulers, samplers, and steps vary depending on the style you aim for and the LoRAs you are using. You should adjust them depending on your task.

🔧 Steps and CFG guidance

• CFG guidance 3.5 – 4 (The higher the value, the more strictly the model follows the prompt. It’s better to keep it at the recommended range.)

• Start from 25+ steps

• Fewer than 20 → risk of pixelation or poor details (depends on sampler/scheduler combo)

• 25–40 → balanced quality/speed (You probably don’t need more than 50 steps)

🪛sampler/scheduler combo

• A quick reminder, if you’re using a LoRA, it’s best to start with the sampler/scheduler combination recommended by the LoRA’s author.
  Otherwise, you may end up with unstable or low-quality results.

  But the most important thing, you should always experiment and find what works best for your own workflow.
  Below are my personally tested combinations that consistently deliver good results with Chroma 👇

  🌀 Res-multistep – Beta

  My go-to option. Works great for both stylized and realistic images.
  In realistic prompts, it produces sharper and more defined details.

  🎨 Euler – Normal / Beta

  Versatile and balanced — performs well for both realistic and stylized prompts.

  📸 DPM++ 2M – sgm_uniform

  Used together with TijuanaSlumlord’s LoRAs , "Balance of speed and quality"

  🔧 DPM_2 – Beta

  Used with TijuanaSlumlord workflows, "slow high quality"

  💡 DDIM – Beta

  Also works well with TijuanaSlumlor

  ⚠️ Note:
  There are more possible combinations of schedulers and samplers, but I haven’t really tested them yet, so I can’t recommend them for now.

4. Prompting Tips

• Chroma is sensitive to prompting. It understands plain English (natural language), you can also rely on the official Black Forest Labs or CivitAI Flux guides as a starting point.

• Chroma supports negative prompts, but they are most effective when used only when needed.

  • If a realism prompt drifts into anime/cartoon style

  • If the output contains unwanted text

• Chroma was partially trained on Danbooru-style tags.
  Writing tags is not strictly necessary, but for me they can help structure prompts and clarify intent.

🔑 Why i Use Tags

• Immediately defines medium (digital_media_(artwork), photography_(artwork), oil_painting_(artwork), sketch, retro_snapshot)

• Helps stabilize output for stylization vs realism

🎨 The “aesthetic” Tag

• If you want to use it, apply the tag in stylized prompts, not realistic ones. I haven’t noticed much difference when using this tag in realism-focused prompts.

• The aesthetic (1–11) tag is can be important:
  👉 Without LoRAs, image quality in Chroma can be random. Using the right aesthetic value gives more consistent results.

  • aesthetic 1-7 → amateur

  • aesthetic 7-10 → professional

  • aesthetic 11 → vibrant, synthetic, can cause AI-slop results

  While aesthetic (1–11) can stabilize style and improve consistency, you should be careful:

  • In complex, long prompts, adding an aesthetic tag can confuse the model and cause:

    • Concept bleeding

    • Lost adherence to detailed descriptions

✒️Power in Simplicity

• Not every super-long, hyper-detailed prompt will give you the best result.
  Often it’s better to write a shorter but well-structured prompt, than a very detailed one overloaded with unnecessary words.

• Example of unnecessary verbosity/ "bad prompt" :

“A cat sitting on a vintage chair, against blue wallpaper with golden flowers that emphasize the aristocratic atmosphere of the room, giving the whole image a noble and majestic vibe, highlighting the refinement of the setting.”

❌ The phrase “aristocratic atmosphere” doesn’t improve the result , it can actually confuse Chroma.

• Example of "good prompt" :

Painting of a cat, sitting on a vintage chair. Behind him, blue wallpaper with golden details.

👉 Short, direct, and much more effective.
⚠️ Note:
Using words that describe the vibe or atmosphere of an image can still be useful ,just don’t overuse them.

4.1 Realism Prompts

• If you want to learn more about realistic image generation, check out the links below, they include settings used by different people and examples of the images they created.
  I also recommend doing your own research to find the workflow that suits you best.
  

Link # 1
Link # 2

• In short, as simple as it sounds, to achieve realistic results you need to clearly describe what you want to see in the final image. You can do this yourself, but in my experience, LLMs can do it better and faster.
  

So you can get straight to the point and simply describe what’s happening in the image, for example:

A stylish young person with short curly hair is enjoying a bowl of ramen noodles. They wear oversized white sunglasses with orange-tinted lenses and a white blazer, exuding a retro yet modern aesthetic. They are holding chopsticks, lifting steaming noodles toward their mouth, while resting their head lazily on one hand. The ramen bowl is ornate with a blue pattern, filled with rich red broth, noodles, vegetables, and pieces of meat. The setting has a sunlit, cinematic vibe with soft golden light casting strong shadows against a tiled wall in the background. The mood is vibrant, fashionable, and slightly playful, like a scene from a trendy editorial photoshoot.

• If you want a realistic image that looks like it was posted on social media, you should mention the media where it was “posted” on (Reddit, Instagram, Twitter, Snapchat, Flickr, OnlyFans).

This is a candid amateur 2010s-era photograph, posted on social media sites such as Reddit, Instagram, Twitter, Snapchat or OnlyFans. The subject is a normal person with natural skin, captued in real life. This is a close-up headshot photograph of a young woman with long hair and a large bust wearing a polka dot string bikini, standing on a windswept beach and illuminated by golden rays of sunshine at dawn.

• If you want the result to look like an amateur selfie, you should start your prompt with something like “A candid self-portrait…”

digital_media_(artwork), absurd_res, surreal, retro_snapshot.

A candid self-portrait of a gothic-inspired young woman holding a vintage Canon PowerShot digital camera, the display screen showing her mirrored pose with a peace sign.

She has long dark hair with bangs, subtle eyeliner, and a confident smirk. Her nails are long with black-and-white Hello Kitty themed nail art, the same stickers decorating the old camera.

She wears a black choker and layered jewelry, adding to the gothic aesthetic.

Background is minimal — a plain white wall and beige curtains, softly illuminated by flat artificial lighting.

The image carries the nostalgic feeling of early 2000s bedroom snapshots, lo-fi casual realism, indie internet culture vibe.

⚠️Realism Prompting Tip

When writing prompts for photorealistic images in Chroma, avoid words like:

• realistic

• hyper-realistic

• ultra-realistic

These terms can confuse the model, producing inconsistent or artificial results.

👉 Instead :

• Just describe what you want to see in the generated image.

• At the very beginning, you can also specify the style of the image by starting your prompt with phrases like:

  
  - This is a candid amateur photograph of...
  - Photo of…
  - Amateur photo of…
  - Professional studio portrait …
  - Retro snapshot…
  

• If your images are still bad - skill issue 😅, I recommend using realism-oriented LoRAs.
  For example, “Chroma – Professional Photos” can fix deformed faces and other visual artifacts, making your image look much more natural and polished.

• Also, try changing your prompting approach, shorten your prompt or rephrase it using different words.
  Sometimes a simpler or more direct phrasing can completely change how Chroma interprets your idea.

4.2 Stylized Prompts

• When creating stylized images (drawings, illustrations, concept art, posters), it’s important to guide Chroma by clearly describing the art direction.

  🔑 Key Rules

1. Define the genre or direction:

   • Concept art, character sheet, comic panel, poster, surrealism, fantasy illustration.

2. Specify the medium / technique:

   • Oil painting, watercolor, acrylic, graphite sketch, pixel art, digital painting.

3. Describe brushwork or texture (optional):

   • “rough brush strokes,” “thin watercolor washes,” “bold ink lines,” “cross-hatching.”

4. Use the aesthetic tag (I wrote more about it above) or try different LoRAs to achieve the desired result.

Example:

aesthetic 11, painterly, digital_media_(artwork), absurd_res, dark_fantasy. A close-up portrait of a young androgynous figure with messy dark hair, pale skin, and piercing eyes glowing faintly in hues of pink and crimson. The face is tilted slightly forward, lips parted as if about to speak, giving a subtle air of tension and mystery. Their attire is a tattered white dress shirt under fragments of dark armor, blending modern and medieval motifs. Loose painterly brushstrokes dominate the piece, layered textures evoking a tactile oil-on-canvas feel. A pixelated red heart hovers faintly near the shoulder, contrasting playfully with the otherwise grim palette. The background dissolves into swirling, abstract shadows and golden flecks, as if the void itself is bleeding light. The overall tone balances raw beauty and haunting surrealism, charged with intensity and quiet menace.

5. Using LoRAs with Chroma

• One of the reasons why I love Chroma is that you can achieve great results even without using any LoRAs.
  However, Chroma is still a neutral base model, which means the results can sometimes be random or inconsistent, especially when you want a specific look or aesthetic.

• So, if your generations don’t fully match your expectations or you want more consistency, you should use LoRAs.
  They help guide the model toward a particular visual style, lighting mood, or realism level.

💡 Quick Tips for Using LoRAs

1. Check creator notes - most LoRAs include recommended CFG, steps, or samplers.

2. Look at author workflows - they often show the exact settings used for best results.

3. Chroma is Flux-based - some Flux LoRAs may work, but you’ll need to experiment to see which ones are compatible.

5.1📸 For Realistic Results

Use realism-oriented LoRAs trained specifically for Chroma or Flux-based models.
Some good examples:

• Chroma – Professional Photos → Magic LoRA that can fix blurry or sloppy images and make them look more realistic (has a strong influence).

• Absolute CINEMA (Chroma) → Makes images feel more cinematic. It has a softer influence on the overall composition, mainly adjusting tone and lighting while preserving the original look of the image.

• Chroma Realistic Slide → The name already explains what it does. There’s also a detailed description available on the author’s page.

There are also a few useful LoRAs for NSFW generation or for creating specific 1980s movie aesthetics, I can’t list them all here, you’ll have to explore and find them yourself.

5.2🎨 For Stylized or Artistic Prompts

If you prefer painterly, anime, or conceptual results, use LoRAs made for stylized work.
Examples:

• Chroma - Random Illustration/Anime Mashup → boosts artistic look, brush texture, and color richness

• Tons of Character and Style LoRAs by Shrekman17 → Huge thanks to him for converting so many LoRAs for Chroma. These cover a wide range of artistic styles and character aesthetics.

• Use the aesthetic tag (I wrote more about it above) or try different LoRAs to achieve the desired result

There are also many other LoRAs for different artistic styles — I’ll probably include some of them in the next version of this guide.

6. Learning Tips

• The Chroma community has many talented creators, including big names with thousands of followers.

• They also create LoRAs that expand Chroma’s abilities.

• Many of their works are regularly featured on the CivitAI homepage (Featured Images section).

👉 The best way to improve is by copying and recreating images you like:

• If an image has metadata, check the workflow:

• See the samplers, schedulers, steps, CFG used.

• Study how they structured their prompts.

• Then try to reproduce or remix their results.

• Also you can feed this guide to ChatGPT (or any LLM you prefer).

Then you can:

• • Ask it to generate prompts based on your description.

  • Ask it to rewrite or optimize existing prompts.

  • Even request a prompt that matches an already existing image.

👉 This way, you’ll train yourself to write better prompts step by step.

Thank You❤️

Thank you for reading my Chroma Guide v0.4!
If you have suggestions, corrections, or tips  I will be very happy to hear them.

👉 Let’s make this guide better together as a community, so Chroma can reach its full potential for everyone.