I have spent more time with HiDream, trying to learn the best practices and approaches to prompting. For HiDream, prompting is everything. You cannot get good results unless you have refined and perfect prompts. Bad prompts will result in poor image quality.
To determine the best practices for prompting with HiDream, I used Gemini Deep Research to generate a report based on all available official documentation and community discussion related to prompting with HiDream's full model.
> Please note that this is only related to the full model, as the full model allows for negative prompts and additional settings like cfg to alter creativity. The full report from Gemini Deep Research is attached to the right-hand side of this article as a PDF attachment, should you wish to read it.
Conclusion from the Report
The following is a direct except from the report:
"The primary strategy should be the use of detailed, descriptive natural language. Prompts should be structured as ONLY!! coherent sentences or paragraphs that clearly articulate the desired outcome. While flexibility exists, organizing the prompt logically can aid interpretation.
A Potential Prompt Flow
Overall Scene/Style/Mood Begin with terms defining the general aesthetic or atmosphere
Examples:
"A photorealistic macro photo...""A surreal, cinematic painting...""An illustration in the style of a vintage travel poster..."
Primary Subject & Action Clearly define the main focus of the image and any actions involved
Examples:
"...of a mechanical bee collecting pollen from a circuit board flower...""...depicting a tiny navy battle inside a kitchen sink...""...showing an astronaut playing chess on the moon..."
Composition & Framing Specify camera angles, shot types, depth of field, or compositional rules
Examples:
"...close-up shot, shallow depth of field...""...viewed through a rainy window pane, rule of thirds composition...""...wide-angle landscape shot..."
Lighting & Atmosphere Describe the lighting conditions and resulting mood
Examples:
"...lit by warm, afternoon sunlight filtering through leaves...""...under dramatic, volumetric studio lighting...""...eerie blue moonlight casting long shadows..."
Best Practices for Negative Prompts
Utility
Negative prompts are likely usable and recommended when working with the HiDream Full model, particularly when employing CFG values greater than 2.0, to refine outputs and steer away from undesired elements.
Content
Due to a lack of specific HiDream Full negative prompt examples in the analyzed materials, users should start with standard negative prompts common for other diffusion models:
"low quality, blurry, deformed, text, watermark, signature, ugly, disfigured, mutation, extra limbs"
Based on potential model tendencies, explicitly negating unwanted content like "NSFW, nudity" might also be necessary in some contexts.
Refine the negative prompt based on observed artifacts or unwanted elements in the generated images. User experimentation is key here."
Avoiding Watermarks & Artifacts
If you read the full report, you’ll notice that I have omitted parts 5 (Key Details & Modifiers) and 6 (Quality/Technical Terms - Optional).
I omitted these because during testing, I found that including them caused a higher likelihood of a PJ watermark or text being generated at the bottom of the image.
Example Issue:
Positive Prompt: "A highly detailed photograph conveying a sense of intricate fragility. The central subject is a life-sized figure of a man meticulously constructed from layered, overlapping strips of vibrant colored paper (think iridescent blues, fiery oranges, and deep purples). His right arm, from the forearm down, is actively unraveling; individual paper ribbons are visibly peeling away from the main form, catching an unseen current and floating gently upwards. The composition is a medium close-up shot, focusing on the man's torso and the dissolving right arm, set against a simple, softly blurred neutral background to maintain focus on the figure. Soft, diffused lighting illuminates the scene, emphasizing the vibrant colors and intricate textures of the layered paper, the feathered edges of the strips, and casting gentle shadows, while highlighting the almost translucent quality of the detached, floating pieces. Key details include the visible construction with countless strip edges, the delicate curling and twisting motion of the ascending strips, and the clear point of dissolution on the arm where the structure gives way."
Negative Prompt: "low quality, blurry, deformed, text, watermark, signature, ugly, disfigured, mutation, extra limbs, worst quality, lowres, title, white title, letters, writing, PJ"Including these sections resulted in unwanted watermarks—removing them maintained image quality while eliminating text artifacts.
Positive Prompt: "A highly detailed photograph conveying a sense of intricate fragility. The central subject is a life-sized figure of a man meticulously constructed from layered, overlapping strips of vibrant colored paper (think iridescent blues, fiery oranges, and deep purples). His right arm, from the forearm down, is actively unraveling; individual paper ribbons are visibly peeling away from the main form, catching an unseen current and floating gently upwards. The composition is a medium close-up shot, focusing on the man's torso and the dissolving right arm, set against a simple, softly blurred neutral background to maintain focus on the figure. Soft, diffused lighting illuminates the scene, emphasizing the vibrant colors and intricate textures of the layered paper, the feathered edges of the strips, and casting gentle shadows, while highlighting the almost translucent quality of the detached, floating pieces."
Negative Prompt: "low quality, blurry, deformed, text, watermark, signature, ugly, disfigured, mutation, extra limbs, worst quality, lowres, title, white title, letters, writing, PJ"I don't know why this happens but not including the key elements section from the report stops it. I know it's easy to remove it in post editing but it's odd that it happens and is clearly triggered either by certain word or phrasing.
Another issue that is mentioned in the report is it suggests avoiding "word packing" at all costs, yet users found on Reddit are doing precisely that by adding phrases like:
"4K, 8K, detailed, sharp focus, high quality"
This word packing should never be used—it increases the chance watermarks and nearly always lowers image quality.
Positive Prompt: "Surreal cinematic painting, colossal ancient stone world turtle, carrying miniature city, overgrown lush civilization, moss, giant roots, drifting through nebula, nebula-filled cosmos, glowing city, subtle life signs, smoke tendrils, tiny lit windows, dramatic low-angle shot, looking up, immense scale, cosmic background, framed by swirling stardust clouds, bioluminescent nebula illumination, ethereal blues, purples, greens light, warm orange city glow, combined lighting, craggy shell texture, detailed, cosmic wonder, miniature scale, 4k, 8k. high quality, best detail."
Negative Prompt: "low quality, blurry, deformed, text, watermark, signature, ugly, disfigured, mutation, extra limbs, worst quality, lowres, title, white title, letters, writing, PJ"
Positive Prompt:"A surreal, cinematic painting depicting a colossal, ancient stone turtle carrying a lush, overgrown city on its back, slowly drifting through a nebula-filled cosmos. The primary subject is the 'world turtle' and the miniature, glowing civilization nestled amongst moss and giant roots upon its shell; subtle signs of life like smoke tendrils from chimneys or tiny lit windows should be visible. The composition is a dramatic low-angle shot, looking up at the immense turtle, emphasizing its scale against the cosmic background, framed by swirling stardust clouds. The scene is illuminated by the bioluminescent glow of the nebula itself – casting ethereal blues, purples, and greens across the turtle's craggy shell and the city structures – combined with the warm, inviting orange glow emanating from the city lights."
Negative Prompt: "low quality, blurry, deformed, text, watermark, signature, ugly, disfigured, mutation, extra limbs, worst quality, lowres, title, white title, letters, writing, PJ"Note on Realism
It's interesting that the report mentions difficulty generating realistic images, as enragedantelope pointed out the same thing in my initial thoughts post earlier.
After testing this new method, you can see it is possible to get realism, as with the mime photo below, but it is incredibly difficult.
Positive Prompt: "Embodying a highly stylized, theatrical aesthetic charged with comedic anticipation, the scene features a classic mime—defined by stark white face paint, precise black accents, pristine gloves, and crisp black-and-white attire—frozen dramatically mid-throw after releasing a projectile. Their body language directs focus towards a plain white paper plate, piled high with fluffy whipped cream, caught perfectly suspended in mid-air just after its release, occupying the space directly before them in the composition. Sharp, theatrical spotlighting isolates the mime and plate against a dark, undefined background, creating high contrasts and a clean, focused atmosphere."
Negative Prompt: "low quality, blurry, deformed, text, watermark, signature, ugly, disfigured, mutation, extra limbs, worst quality, lowres, title, white title, letters, writing, PJ"While this image did not follow the prompt precisely (one of the few instances where it didn't), that image likely gained its realistic look because it’s a specific niche style that doesn't often get photoshopped or edited.
Trying to get an image of a woman to appear that realistic is something I have not achieved yet.
I think EnragedAntelope is correct—HiDream would need a LoRA training specifically for realism.
Final Thoughts
From my testing, the majority of my opinions remain the same:
HiDream is a great model, and its MIT license is valuable, allowing people to sell what they create.
A lot of users have avoided Flux due to its restrictive license.
Having an open-source model as high quality as Flux is something we should all embrace.
Encouraging skilled developers to build tools and work on getting better outputs, like what happened with ipAdapters and ControlNets in SDXL, is crucial. We only achieve that if we start showing an interest in the model.
> Further information from the report is available in the attachments section to the right of this page. It's 11 pages long and quite boring, but provides insights into adjusting CFG, recommending: Start with CFG 3 Go up to 5 if it's not following the prompt, amongst other tips.
Here is a couple more images for good measure, some of these I stole the prompt from the top flux image board and then altered them using this method. I assume those who post the prompt with their image are fine with remixes, contact me if not. :
credit to original: https://civitai.com/images/64399178
Positive prompt: a high-quality rendering reminiscent of classic 1940s film noir, conveying mystery and duality. The primary subject is a person's face, visually split: the left half is yellow, covered in intricate black and blue symbols/doodles, while the right half is dark, revealing underlying mechanical components, with a bright yellow left eye standing out. A hood partially obscures the face. Use an intense close-up view, tightly framed on the face to emphasize these details. Employ dramatic, high-contrast lighting typical of film noir to highlight the textures and the stark difference between the two halves, enhancing the mysterious atmosphere.
Negative Prompt: "low quality, blurry, deformed, text, watermark, signature, ugly, disfigured, mutation, extra limbs, worst quality, lowres, title, white title, letters, writing, PJ"
Positive Prompt: "A richly detailed digital painting evoking quiet intelligence and ancient alien mystery. The central focus is a tall, slender insectoid alien, reminiscent of a praying mantis with an iridescent exoskeleton (shifting greens and purples), standing within a vast, dimly lit alien archive. The alien is captured mid-action, carefully manipulating and examining a complex, brightly glowing holographic star chart projected from a small crystalline device held in its delicate, multi-jointed upper limbs. The composition is a close-up shot, concentrating on the alien's upper body, its large, multifaceted black eyes reflecting the hologram's light, and its intricate interaction with the projection. Soft, ambient light filters down from high above, hinting at the immense scale of the archive, while the primary illumination comes from the cool, blue-white volumetric glow of the star chart, casting intricate shadows across the alien's form and the nearby towering shelves filled with glowing data crystals that recede into deep shadow, creating significant depth."
Negative Prompt: "low quality, blurry, deformed, text, watermark, signature, ugly, disfigured, mutation, extra limbs, worst quality, lowres, title, white title, letters, writing, PJ"
credit to original: https://civitai.com/images/27763004
Positive Prompt: "a detailed digital illustration in a semi-realistic, vibrant, high-contrast style influenced by 90s retro-futuristic anime, evoking a dystopian sci-fi feel. The subject is a serious-looking woman with dark, shoulder-length hair and green eyes, possessing a slender, athletic cyborg body (mechanical arms/chest) clad in a vivid red, high-tech futuristic suit (military/space gear style). Capture her in a dynamic pose. Use a medium or full shot to show the character's pose within the context of a dark, abandoned industrial background featuring greenish tones, metallic pipes, and machinery. Apply a cool color palette with a slightly gritty texture, using lighting that emphasizes the contrast between her red suit and the dark environment, creating a tense, dystopian atmosphere."
Negative Prompt: "low quality, blurry, deformed, text, watermark, signature, ugly, disfigured, mutation, extra limbs, worst quality, lowres, title, white title, letters, writing, PJ"
Credit to the original: https://civitai.com/images/37410111
Positive Prompt: "a dark fantasy horror scene embodying nightmares, curses, and malevolent magic. Depict a dark, hooded figure standing within dense fog. Its eyes should glow like fire, and a malicious yellow skull face must be visible beneath the dark purple hood. Show a clawed hand reaching forward, emitting swirling, smoke-like magical energy. Use a medium shot focusing on the figure's menacing upper body and outstretched hand, partially obscured by thick fog to enhance mystery. The lighting should be dim and fog-drenched, with the primary light sources being the glowing eyes and magical energy, creating a threatening, poisonous, terrifying, and ancient atmosphere."
Negative Prompt: "low quality, blurry, deformed, text, watermark, signature, ugly, disfigured, mutation, extra limbs, worst quality, lowres, title, white title, letters, writing, PJ"This is the structure I will be following while using HiDream going forward, you can see my account profile to see what I post in the future. I usually include prompts in my own stuff.