Product Manager - B2B & Enterprise

Melius raises $20M after ex-ramp founders scrapped first AI ad product. New York-based AI startup Melius has closed a $25 million funding round, comprising a $20 million Series A led by CRV and a $5 million seed round backed by General Catalyst. The company, co-founded by former Ramp engineers Joowon Kim, Young Kim, and Arnav Ramu, announced the investment following a July stealth launch that has already generated over $1 million in annualized recurring revenue. Melius operates in the expanding artificial intelligence marketing sector but emerged from a decisive strategic pivot. Initially, the founding team developed an AI-driven performance marketing tool designed to help advertisers manage and optimize budget allocation. After six months of development, the co-founders determined the original concept lacked sufficient market traction and subsequently dismantled the existing codebase. The team redirected its engineering efforts toward building an AI-powered creative generation platform, now positioned as an agents lab for end-to-end campaign and asset creation. The current platform enables users to generate advertisements, images, and video content through natural language prompts. Co-founder Joowon Kim, who has long expressed a personal interest in digital media production, positioned the tool as an accessible solution that allows both novice creators and experienced creative directors to translate conceptual ideas into polished marketing assets without specialized technical software. Melius enters a competitive landscape that includes high-growth rivals such as Higgsfield, which achieved a $5.4 billion valuation and over $700 million in annualized revenue, alongside emerging players like Krea and Flora AI. Despite the crowded market, Kim maintains that the total addressable space for AI-generated creative content remains sufficiently large to support multiple industry leaders. The presence of numerous entrants signals strong enterprise demand and validates the commercial viability of automated creative workflows. The funding round underscores investor confidence in Melius's pivot strategy and its ability to capture early market traction. By transitioning from ad spend optimization to generative creative infrastructure, the startup positions itself at the intersection of artificial intelligence and digital marketing production. As traditional advertising workflows continue to digitize, Melius aims to streamline the creative pipeline, offering brands and agencies a scalable alternative to conventional content development cycles. This news is intelligently aggregated by AI to deliver industry updates efficiently. It does not constitute opinions or advice.

Higgsfield brings Ideogram 4.5 to stop repeated edits from destroying images. Higgsfield has added Ideogram 4.5 to its platform, bringing the drift-free multi-turn edit model to a new audience with native 2K output. · 5 hrs ago Read 5 min * Ideogram 4.5 is now live on Higgsfield, Ideogram's site, and the API. * Core promise: multi-turn edits without pixel drift, color shift, or artifact buildup. * Four quality tiers, 0.8¢ to 22¢ per image, all native 2K output. * Supports up to 4 reference images, optional masks, and crop-then-stitch high-res editing. * Ranks #18 overall on Image Edit Arena with 1351 points across categories. * Open weights promised soon, following Ideogram 4.0's Apache 2.0 release. Repeated generative edits can degrade untouched areas as pixel shifts, color changes, and artifacts accumulate. Higgsfield now offers Ideogram 4.5, a model designed to preserve the source image across successive edits. On Higgsfield's model page, users can target text, products, colors, lighting, or another isolated detail at the source resolution. Ideogram calls 4.5 its "most precise edit model" in the launch post and says its high-precision mode restores unchanged pixels exactly. That preservation claim concerns decoded pixel values outside the edited region. File metadata, compression, and container bytes can still differ between the input and output. Why repeated edits decay. Many generative editors resynthesize more of an image than the requested region. Small variations appear outside the target area, then become part of the input for the next pass. A sequence of otherwise successful edits can therefore alter faces, typography, product geometry, and color balance. Ideogram's published comparisons show GPT Image and Nano Banana accumulating visible artifacts after several turns, with 4.5 retaining a cleaner source image. These examples come from the vendor, so production teams should verify the result with their own images, masks, and edit sequences. Crop-level editing also becomes easier when dimensions and surrounding pixels remain stable. A client can crop a billboard from a 2K scene, replace its copy, and composite the crop back into the original without resampling the full image. API routes and controls. Developers can access the model through Ideogram's site and official API, Higgsfield, and partners such as the fal model page. Provider schemas and billing systems may differ, but Ideogram's API exposes the following controls: | Control | Behavior | | Workflows | Precise Edit returns the source dimensions. Generate + Edit combines generation with an edit request. | | Reference images | Up to four references are supported, reduced to three when the request includes a mask. | | edit_precision | regular is the default. high invokes Precise Edit and restores unchanged source pixels. | | Quality | Regular precision supports tiers from very_low through high. High precision supports low through very_high. | | Output size | The output defaults to the source dimensions. Ideogram advertises native 2K processing. | | Prompt length | Prompts can contain up to 10,000 characters. | At launch, Ideogram listed four billed quality modes ranging from 0.8¢ to 22¢ per image, equivalent to $0.008 to $0.22. Higgsfield and fal may charge through separate credit or pricing systems, so API cost comparisons should use each provider's current rates. Typography, products, and restoration. Ideogram has built its product around reliable in-image typography, and 4.5 extends that focus to existing designs. The model can replace or translate stylized text while retaining layout, color, and surrounding artwork. * Product variants: Recolor an object or change scene lighting while updating its shadows, reflections, and highlights. * Campaign localization: Translate packaging, posters, or advertising copy without rebuilding the full composition. * Photo restoration: Remove scratches first, add color in a later pass, then apply enhancement as a separate edit. * Interior design: Swap furniture, finishes, and palettes while preserving the room's architecture. Arena results temper the pitch. The Image Edit Arena provides a broader measure based on blind user preferences across general editing tasks. At the cited leaderboard snapshot, Ideogram 4.5 scored 1,351 points and ranked 18th overall. | Category | Rank | | 3D Imaging & Modeling | 14 | | Product & Commercial Design | 15 | | Cartoon, Anime & Fantasy | 17 | | Photorealistic & Cinematic | 17 | | Text Rendering | 17 | | Portraits | 18 | Those rankings place 4.5 in the top 20 without leading a measured category. The leaderboard covers general editing preferences, whereas Ideogram's main claim centers on pixel preservation during iterative work. Arena ratings are live and can move as new votes and models arrive. Open weights remain pending. Ideogram's launch post says open weights are coming "soon," but the company has not provided a release date, license, model size, or hardware requirements. Hosted services remain the available deployment path until those details arrive. Structural edits need more turns. Ideogram recommends splitting large transformations into bounded steps, such as replacing copy, changing a product color, and adjusting lighting in separate requests. A single prompt that restructures an entire scene falls outside the model's strongest workflow, and generation-focused models may produce better results for broad composition changes. Test the preservation claim. A production evaluation should measure pixel stability alongside visual quality, latency, and cost. A compact test plan can cover the main failure modes: * Decode the input and output into pixel arrays, then count changed pixels outside the requested mask or region. * Run five to ten sequential edits and track color drift, geometry changes, and artifact accumulation after each turn. * Verify output dimensions, alpha handling, color profiles, metadata, and lossless export behavior. * Test small text, stylized lettering, multilingual copy, reflections, shadows, faces, and product logos. * Compare provider-specific latency, rate limits, moderation rules, credit usage, and retry behavior. Iterative product photography, localized campaigns, typography changes, restoration, and interior variants are the clearest use cases for Ideogram 4.5. Higgsfield makes the model available through its existing image interface, while API users should map provider-specific fields and output handling before replacing an existing model.

AI video startup Higgsfield AI created the 95-minute film Hell Grind in under three weeks with a $500,000 budget, debuting it at Cannes this spring. The company raised $400 million in a Series B round in August at a $5.4 billion valuation, following an $80 million raise eight months earlier. Rival startup Runway raised $315 million in a Series E at a $5.3 billion valuation. The company partnered with Lionsgate and has evolved its technology from restyling existing videos to generating new content from text prompts. Luma secured $900 million in November 2025 from investors including Saudi AI company Humain, Nvidia, and AMD. McKinsey projects AI could influence up to 20% of original content spending on films and TV by 2030. SAG-AFTRA recently ratified a four-year contract with major studios establishing guardrails around AI usage in production.